1 //===-- AArch64ISelLowering.cpp - AArch64 DAG Lowering Implementation ----===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements the AArch64TargetLowering class. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "AArch64ISelLowering.h" 14 #include "AArch64CallingConvention.h" 15 #include "AArch64ExpandImm.h" 16 #include "AArch64MachineFunctionInfo.h" 17 #include "AArch64PerfectShuffle.h" 18 #include "AArch64RegisterInfo.h" 19 #include "AArch64Subtarget.h" 20 #include "MCTargetDesc/AArch64AddressingModes.h" 21 #include "Utils/AArch64BaseInfo.h" 22 #include "llvm/ADT/APFloat.h" 23 #include "llvm/ADT/APInt.h" 24 #include "llvm/ADT/ArrayRef.h" 25 #include "llvm/ADT/STLExtras.h" 26 #include "llvm/ADT/SmallSet.h" 27 #include "llvm/ADT/SmallVector.h" 28 #include "llvm/ADT/Statistic.h" 29 #include "llvm/ADT/StringRef.h" 30 #include "llvm/ADT/Triple.h" 31 #include "llvm/ADT/Twine.h" 32 #include "llvm/Analysis/ObjCARCUtil.h" 33 #include "llvm/Analysis/VectorUtils.h" 34 #include "llvm/CodeGen/CallingConvLower.h" 35 #include "llvm/CodeGen/MachineBasicBlock.h" 36 #include "llvm/CodeGen/MachineFrameInfo.h" 37 #include "llvm/CodeGen/MachineFunction.h" 38 #include "llvm/CodeGen/MachineInstr.h" 39 #include "llvm/CodeGen/MachineInstrBuilder.h" 40 #include "llvm/CodeGen/MachineMemOperand.h" 41 #include "llvm/CodeGen/MachineRegisterInfo.h" 42 #include "llvm/CodeGen/RuntimeLibcalls.h" 43 #include "llvm/CodeGen/SelectionDAG.h" 44 #include "llvm/CodeGen/SelectionDAGNodes.h" 45 #include "llvm/CodeGen/TargetCallingConv.h" 46 #include "llvm/CodeGen/TargetInstrInfo.h" 47 #include "llvm/CodeGen/ValueTypes.h" 48 #include "llvm/IR/Attributes.h" 49 #include "llvm/IR/Constants.h" 50 #include "llvm/IR/DataLayout.h" 51 #include "llvm/IR/DebugLoc.h" 52 #include "llvm/IR/DerivedTypes.h" 53 #include "llvm/IR/Function.h" 54 #include "llvm/IR/GetElementPtrTypeIterator.h" 55 #include "llvm/IR/GlobalValue.h" 56 #include "llvm/IR/IRBuilder.h" 57 #include "llvm/IR/Instruction.h" 58 #include "llvm/IR/Instructions.h" 59 #include "llvm/IR/IntrinsicInst.h" 60 #include "llvm/IR/Intrinsics.h" 61 #include "llvm/IR/IntrinsicsAArch64.h" 62 #include "llvm/IR/Module.h" 63 #include "llvm/IR/OperandTraits.h" 64 #include "llvm/IR/PatternMatch.h" 65 #include "llvm/IR/Type.h" 66 #include "llvm/IR/Use.h" 67 #include "llvm/IR/Value.h" 68 #include "llvm/MC/MCRegisterInfo.h" 69 #include "llvm/Support/Casting.h" 70 #include "llvm/Support/CodeGen.h" 71 #include "llvm/Support/CommandLine.h" 72 #include "llvm/Support/Compiler.h" 73 #include "llvm/Support/Debug.h" 74 #include "llvm/Support/ErrorHandling.h" 75 #include "llvm/Support/KnownBits.h" 76 #include "llvm/Support/MachineValueType.h" 77 #include "llvm/Support/MathExtras.h" 78 #include "llvm/Support/raw_ostream.h" 79 #include "llvm/Target/TargetMachine.h" 80 #include "llvm/Target/TargetOptions.h" 81 #include <algorithm> 82 #include <bitset> 83 #include <cassert> 84 #include <cctype> 85 #include <cstdint> 86 #include <cstdlib> 87 #include <iterator> 88 #include <limits> 89 #include <tuple> 90 #include <utility> 91 #include <vector> 92 93 using namespace llvm; 94 using namespace llvm::PatternMatch; 95 96 #define DEBUG_TYPE "aarch64-lower" 97 98 STATISTIC(NumTailCalls, "Number of tail calls"); 99 STATISTIC(NumShiftInserts, "Number of vector shift inserts"); 100 STATISTIC(NumOptimizedImms, "Number of times immediates were optimized"); 101 102 // FIXME: The necessary dtprel relocations don't seem to be supported 103 // well in the GNU bfd and gold linkers at the moment. Therefore, by 104 // default, for now, fall back to GeneralDynamic code generation. 105 cl::opt<bool> EnableAArch64ELFLocalDynamicTLSGeneration( 106 "aarch64-elf-ldtls-generation", cl::Hidden, 107 cl::desc("Allow AArch64 Local Dynamic TLS code generation"), 108 cl::init(false)); 109 110 static cl::opt<bool> 111 EnableOptimizeLogicalImm("aarch64-enable-logical-imm", cl::Hidden, 112 cl::desc("Enable AArch64 logical imm instruction " 113 "optimization"), 114 cl::init(true)); 115 116 // Temporary option added for the purpose of testing functionality added 117 // to DAGCombiner.cpp in D92230. It is expected that this can be removed 118 // in future when both implementations will be based off MGATHER rather 119 // than the GLD1 nodes added for the SVE gather load intrinsics. 120 static cl::opt<bool> 121 EnableCombineMGatherIntrinsics("aarch64-enable-mgather-combine", cl::Hidden, 122 cl::desc("Combine extends of AArch64 masked " 123 "gather intrinsics"), 124 cl::init(true)); 125 126 /// Value type used for condition codes. 127 static const MVT MVT_CC = MVT::i32; 128 129 static inline EVT getPackedSVEVectorVT(EVT VT) { 130 switch (VT.getSimpleVT().SimpleTy) { 131 default: 132 llvm_unreachable("unexpected element type for vector"); 133 case MVT::i8: 134 return MVT::nxv16i8; 135 case MVT::i16: 136 return MVT::nxv8i16; 137 case MVT::i32: 138 return MVT::nxv4i32; 139 case MVT::i64: 140 return MVT::nxv2i64; 141 case MVT::f16: 142 return MVT::nxv8f16; 143 case MVT::f32: 144 return MVT::nxv4f32; 145 case MVT::f64: 146 return MVT::nxv2f64; 147 case MVT::bf16: 148 return MVT::nxv8bf16; 149 } 150 } 151 152 // NOTE: Currently there's only a need to return integer vector types. If this 153 // changes then just add an extra "type" parameter. 154 static inline EVT getPackedSVEVectorVT(ElementCount EC) { 155 switch (EC.getKnownMinValue()) { 156 default: 157 llvm_unreachable("unexpected element count for vector"); 158 case 16: 159 return MVT::nxv16i8; 160 case 8: 161 return MVT::nxv8i16; 162 case 4: 163 return MVT::nxv4i32; 164 case 2: 165 return MVT::nxv2i64; 166 } 167 } 168 169 static inline EVT getPromotedVTForPredicate(EVT VT) { 170 assert(VT.isScalableVector() && (VT.getVectorElementType() == MVT::i1) && 171 "Expected scalable predicate vector type!"); 172 switch (VT.getVectorMinNumElements()) { 173 default: 174 llvm_unreachable("unexpected element count for vector"); 175 case 2: 176 return MVT::nxv2i64; 177 case 4: 178 return MVT::nxv4i32; 179 case 8: 180 return MVT::nxv8i16; 181 case 16: 182 return MVT::nxv16i8; 183 } 184 } 185 186 /// Returns true if VT's elements occupy the lowest bit positions of its 187 /// associated register class without any intervening space. 188 /// 189 /// For example, nxv2f16, nxv4f16 and nxv8f16 are legal types that belong to the 190 /// same register class, but only nxv8f16 can be treated as a packed vector. 191 static inline bool isPackedVectorType(EVT VT, SelectionDAG &DAG) { 192 assert(VT.isVector() && DAG.getTargetLoweringInfo().isTypeLegal(VT) && 193 "Expected legal vector type!"); 194 return VT.isFixedLengthVector() || 195 VT.getSizeInBits().getKnownMinSize() == AArch64::SVEBitsPerBlock; 196 } 197 198 // Returns true for ####_MERGE_PASSTHRU opcodes, whose operands have a leading 199 // predicate and end with a passthru value matching the result type. 200 static bool isMergePassthruOpcode(unsigned Opc) { 201 switch (Opc) { 202 default: 203 return false; 204 case AArch64ISD::BITREVERSE_MERGE_PASSTHRU: 205 case AArch64ISD::BSWAP_MERGE_PASSTHRU: 206 case AArch64ISD::CTLZ_MERGE_PASSTHRU: 207 case AArch64ISD::CTPOP_MERGE_PASSTHRU: 208 case AArch64ISD::DUP_MERGE_PASSTHRU: 209 case AArch64ISD::ABS_MERGE_PASSTHRU: 210 case AArch64ISD::NEG_MERGE_PASSTHRU: 211 case AArch64ISD::FNEG_MERGE_PASSTHRU: 212 case AArch64ISD::SIGN_EXTEND_INREG_MERGE_PASSTHRU: 213 case AArch64ISD::ZERO_EXTEND_INREG_MERGE_PASSTHRU: 214 case AArch64ISD::FCEIL_MERGE_PASSTHRU: 215 case AArch64ISD::FFLOOR_MERGE_PASSTHRU: 216 case AArch64ISD::FNEARBYINT_MERGE_PASSTHRU: 217 case AArch64ISD::FRINT_MERGE_PASSTHRU: 218 case AArch64ISD::FROUND_MERGE_PASSTHRU: 219 case AArch64ISD::FROUNDEVEN_MERGE_PASSTHRU: 220 case AArch64ISD::FTRUNC_MERGE_PASSTHRU: 221 case AArch64ISD::FP_ROUND_MERGE_PASSTHRU: 222 case AArch64ISD::FP_EXTEND_MERGE_PASSTHRU: 223 case AArch64ISD::SINT_TO_FP_MERGE_PASSTHRU: 224 case AArch64ISD::UINT_TO_FP_MERGE_PASSTHRU: 225 case AArch64ISD::FCVTZU_MERGE_PASSTHRU: 226 case AArch64ISD::FCVTZS_MERGE_PASSTHRU: 227 case AArch64ISD::FSQRT_MERGE_PASSTHRU: 228 case AArch64ISD::FRECPX_MERGE_PASSTHRU: 229 case AArch64ISD::FABS_MERGE_PASSTHRU: 230 return true; 231 } 232 } 233 234 AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM, 235 const AArch64Subtarget &STI) 236 : TargetLowering(TM), Subtarget(&STI) { 237 // AArch64 doesn't have comparisons which set GPRs or setcc instructions, so 238 // we have to make something up. Arbitrarily, choose ZeroOrOne. 239 setBooleanContents(ZeroOrOneBooleanContent); 240 // When comparing vectors the result sets the different elements in the 241 // vector to all-one or all-zero. 242 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent); 243 244 // Set up the register classes. 245 addRegisterClass(MVT::i32, &AArch64::GPR32allRegClass); 246 addRegisterClass(MVT::i64, &AArch64::GPR64allRegClass); 247 248 if (Subtarget->hasFPARMv8()) { 249 addRegisterClass(MVT::f16, &AArch64::FPR16RegClass); 250 addRegisterClass(MVT::bf16, &AArch64::FPR16RegClass); 251 addRegisterClass(MVT::f32, &AArch64::FPR32RegClass); 252 addRegisterClass(MVT::f64, &AArch64::FPR64RegClass); 253 addRegisterClass(MVT::f128, &AArch64::FPR128RegClass); 254 } 255 256 if (Subtarget->hasNEON()) { 257 addRegisterClass(MVT::v16i8, &AArch64::FPR8RegClass); 258 addRegisterClass(MVT::v8i16, &AArch64::FPR16RegClass); 259 // Someone set us up the NEON. 260 addDRTypeForNEON(MVT::v2f32); 261 addDRTypeForNEON(MVT::v8i8); 262 addDRTypeForNEON(MVT::v4i16); 263 addDRTypeForNEON(MVT::v2i32); 264 addDRTypeForNEON(MVT::v1i64); 265 addDRTypeForNEON(MVT::v1f64); 266 addDRTypeForNEON(MVT::v4f16); 267 if (Subtarget->hasBF16()) 268 addDRTypeForNEON(MVT::v4bf16); 269 270 addQRTypeForNEON(MVT::v4f32); 271 addQRTypeForNEON(MVT::v2f64); 272 addQRTypeForNEON(MVT::v16i8); 273 addQRTypeForNEON(MVT::v8i16); 274 addQRTypeForNEON(MVT::v4i32); 275 addQRTypeForNEON(MVT::v2i64); 276 addQRTypeForNEON(MVT::v8f16); 277 if (Subtarget->hasBF16()) 278 addQRTypeForNEON(MVT::v8bf16); 279 } 280 281 if (Subtarget->hasSVE()) { 282 // Add legal sve predicate types 283 addRegisterClass(MVT::nxv2i1, &AArch64::PPRRegClass); 284 addRegisterClass(MVT::nxv4i1, &AArch64::PPRRegClass); 285 addRegisterClass(MVT::nxv8i1, &AArch64::PPRRegClass); 286 addRegisterClass(MVT::nxv16i1, &AArch64::PPRRegClass); 287 288 // Add legal sve data types 289 addRegisterClass(MVT::nxv16i8, &AArch64::ZPRRegClass); 290 addRegisterClass(MVT::nxv8i16, &AArch64::ZPRRegClass); 291 addRegisterClass(MVT::nxv4i32, &AArch64::ZPRRegClass); 292 addRegisterClass(MVT::nxv2i64, &AArch64::ZPRRegClass); 293 294 addRegisterClass(MVT::nxv2f16, &AArch64::ZPRRegClass); 295 addRegisterClass(MVT::nxv4f16, &AArch64::ZPRRegClass); 296 addRegisterClass(MVT::nxv8f16, &AArch64::ZPRRegClass); 297 addRegisterClass(MVT::nxv2f32, &AArch64::ZPRRegClass); 298 addRegisterClass(MVT::nxv4f32, &AArch64::ZPRRegClass); 299 addRegisterClass(MVT::nxv2f64, &AArch64::ZPRRegClass); 300 301 if (Subtarget->hasBF16()) { 302 addRegisterClass(MVT::nxv2bf16, &AArch64::ZPRRegClass); 303 addRegisterClass(MVT::nxv4bf16, &AArch64::ZPRRegClass); 304 addRegisterClass(MVT::nxv8bf16, &AArch64::ZPRRegClass); 305 } 306 307 if (Subtarget->useSVEForFixedLengthVectors()) { 308 for (MVT VT : MVT::integer_fixedlen_vector_valuetypes()) 309 if (useSVEForFixedLengthVectorVT(VT)) 310 addRegisterClass(VT, &AArch64::ZPRRegClass); 311 312 for (MVT VT : MVT::fp_fixedlen_vector_valuetypes()) 313 if (useSVEForFixedLengthVectorVT(VT)) 314 addRegisterClass(VT, &AArch64::ZPRRegClass); 315 } 316 317 for (auto VT : { MVT::nxv16i8, MVT::nxv8i16, MVT::nxv4i32, MVT::nxv2i64 }) { 318 setOperationAction(ISD::SADDSAT, VT, Legal); 319 setOperationAction(ISD::UADDSAT, VT, Legal); 320 setOperationAction(ISD::SSUBSAT, VT, Legal); 321 setOperationAction(ISD::USUBSAT, VT, Legal); 322 setOperationAction(ISD::UREM, VT, Expand); 323 setOperationAction(ISD::SREM, VT, Expand); 324 setOperationAction(ISD::SDIVREM, VT, Expand); 325 setOperationAction(ISD::UDIVREM, VT, Expand); 326 } 327 328 for (auto VT : 329 { MVT::nxv2i8, MVT::nxv2i16, MVT::nxv2i32, MVT::nxv2i64, MVT::nxv4i8, 330 MVT::nxv4i16, MVT::nxv4i32, MVT::nxv8i8, MVT::nxv8i16 }) 331 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Legal); 332 333 for (auto VT : 334 { MVT::nxv2f16, MVT::nxv4f16, MVT::nxv8f16, MVT::nxv2f32, MVT::nxv4f32, 335 MVT::nxv2f64 }) { 336 setCondCodeAction(ISD::SETO, VT, Expand); 337 setCondCodeAction(ISD::SETOLT, VT, Expand); 338 setCondCodeAction(ISD::SETLT, VT, Expand); 339 setCondCodeAction(ISD::SETOLE, VT, Expand); 340 setCondCodeAction(ISD::SETLE, VT, Expand); 341 setCondCodeAction(ISD::SETULT, VT, Expand); 342 setCondCodeAction(ISD::SETULE, VT, Expand); 343 setCondCodeAction(ISD::SETUGE, VT, Expand); 344 setCondCodeAction(ISD::SETUGT, VT, Expand); 345 setCondCodeAction(ISD::SETUEQ, VT, Expand); 346 setCondCodeAction(ISD::SETUNE, VT, Expand); 347 348 setOperationAction(ISD::FREM, VT, Expand); 349 setOperationAction(ISD::FPOW, VT, Expand); 350 setOperationAction(ISD::FPOWI, VT, Expand); 351 setOperationAction(ISD::FCOS, VT, Expand); 352 setOperationAction(ISD::FSIN, VT, Expand); 353 setOperationAction(ISD::FSINCOS, VT, Expand); 354 setOperationAction(ISD::FEXP, VT, Expand); 355 setOperationAction(ISD::FEXP2, VT, Expand); 356 setOperationAction(ISD::FLOG, VT, Expand); 357 setOperationAction(ISD::FLOG2, VT, Expand); 358 setOperationAction(ISD::FLOG10, VT, Expand); 359 } 360 } 361 362 // Compute derived properties from the register classes 363 computeRegisterProperties(Subtarget->getRegisterInfo()); 364 365 // Provide all sorts of operation actions 366 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom); 367 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom); 368 setOperationAction(ISD::SETCC, MVT::i32, Custom); 369 setOperationAction(ISD::SETCC, MVT::i64, Custom); 370 setOperationAction(ISD::SETCC, MVT::f16, Custom); 371 setOperationAction(ISD::SETCC, MVT::f32, Custom); 372 setOperationAction(ISD::SETCC, MVT::f64, Custom); 373 setOperationAction(ISD::STRICT_FSETCC, MVT::f16, Custom); 374 setOperationAction(ISD::STRICT_FSETCC, MVT::f32, Custom); 375 setOperationAction(ISD::STRICT_FSETCC, MVT::f64, Custom); 376 setOperationAction(ISD::STRICT_FSETCCS, MVT::f16, Custom); 377 setOperationAction(ISD::STRICT_FSETCCS, MVT::f32, Custom); 378 setOperationAction(ISD::STRICT_FSETCCS, MVT::f64, Custom); 379 setOperationAction(ISD::BITREVERSE, MVT::i32, Legal); 380 setOperationAction(ISD::BITREVERSE, MVT::i64, Legal); 381 setOperationAction(ISD::BRCOND, MVT::Other, Expand); 382 setOperationAction(ISD::BR_CC, MVT::i32, Custom); 383 setOperationAction(ISD::BR_CC, MVT::i64, Custom); 384 setOperationAction(ISD::BR_CC, MVT::f16, Custom); 385 setOperationAction(ISD::BR_CC, MVT::f32, Custom); 386 setOperationAction(ISD::BR_CC, MVT::f64, Custom); 387 setOperationAction(ISD::SELECT, MVT::i32, Custom); 388 setOperationAction(ISD::SELECT, MVT::i64, Custom); 389 setOperationAction(ISD::SELECT, MVT::f16, Custom); 390 setOperationAction(ISD::SELECT, MVT::f32, Custom); 391 setOperationAction(ISD::SELECT, MVT::f64, Custom); 392 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom); 393 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom); 394 setOperationAction(ISD::SELECT_CC, MVT::f16, Custom); 395 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom); 396 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom); 397 setOperationAction(ISD::BR_JT, MVT::Other, Custom); 398 setOperationAction(ISD::JumpTable, MVT::i64, Custom); 399 400 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom); 401 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom); 402 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom); 403 404 setOperationAction(ISD::FREM, MVT::f32, Expand); 405 setOperationAction(ISD::FREM, MVT::f64, Expand); 406 setOperationAction(ISD::FREM, MVT::f80, Expand); 407 408 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand); 409 410 // Custom lowering hooks are needed for XOR 411 // to fold it into CSINC/CSINV. 412 setOperationAction(ISD::XOR, MVT::i32, Custom); 413 setOperationAction(ISD::XOR, MVT::i64, Custom); 414 415 // Virtually no operation on f128 is legal, but LLVM can't expand them when 416 // there's a valid register class, so we need custom operations in most cases. 417 setOperationAction(ISD::FABS, MVT::f128, Expand); 418 setOperationAction(ISD::FADD, MVT::f128, LibCall); 419 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand); 420 setOperationAction(ISD::FCOS, MVT::f128, Expand); 421 setOperationAction(ISD::FDIV, MVT::f128, LibCall); 422 setOperationAction(ISD::FMA, MVT::f128, Expand); 423 setOperationAction(ISD::FMUL, MVT::f128, LibCall); 424 setOperationAction(ISD::FNEG, MVT::f128, Expand); 425 setOperationAction(ISD::FPOW, MVT::f128, Expand); 426 setOperationAction(ISD::FREM, MVT::f128, Expand); 427 setOperationAction(ISD::FRINT, MVT::f128, Expand); 428 setOperationAction(ISD::FSIN, MVT::f128, Expand); 429 setOperationAction(ISD::FSINCOS, MVT::f128, Expand); 430 setOperationAction(ISD::FSQRT, MVT::f128, Expand); 431 setOperationAction(ISD::FSUB, MVT::f128, LibCall); 432 setOperationAction(ISD::FTRUNC, MVT::f128, Expand); 433 setOperationAction(ISD::SETCC, MVT::f128, Custom); 434 setOperationAction(ISD::STRICT_FSETCC, MVT::f128, Custom); 435 setOperationAction(ISD::STRICT_FSETCCS, MVT::f128, Custom); 436 setOperationAction(ISD::BR_CC, MVT::f128, Custom); 437 setOperationAction(ISD::SELECT, MVT::f128, Custom); 438 setOperationAction(ISD::SELECT_CC, MVT::f128, Custom); 439 setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom); 440 441 // Lowering for many of the conversions is actually specified by the non-f128 442 // type. The LowerXXX function will be trivial when f128 isn't involved. 443 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom); 444 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom); 445 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom); 446 setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i32, Custom); 447 setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i64, Custom); 448 setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i128, Custom); 449 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom); 450 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom); 451 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom); 452 setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i32, Custom); 453 setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i64, Custom); 454 setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i128, Custom); 455 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom); 456 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom); 457 setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom); 458 setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i32, Custom); 459 setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i64, Custom); 460 setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i128, Custom); 461 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom); 462 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom); 463 setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom); 464 setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i32, Custom); 465 setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i64, Custom); 466 setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i128, Custom); 467 setOperationAction(ISD::FP_ROUND, MVT::f16, Custom); 468 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom); 469 setOperationAction(ISD::FP_ROUND, MVT::f64, Custom); 470 setOperationAction(ISD::STRICT_FP_ROUND, MVT::f16, Custom); 471 setOperationAction(ISD::STRICT_FP_ROUND, MVT::f32, Custom); 472 setOperationAction(ISD::STRICT_FP_ROUND, MVT::f64, Custom); 473 474 setOperationAction(ISD::FP_TO_UINT_SAT, MVT::i32, Custom); 475 setOperationAction(ISD::FP_TO_UINT_SAT, MVT::i64, Custom); 476 setOperationAction(ISD::FP_TO_SINT_SAT, MVT::i32, Custom); 477 setOperationAction(ISD::FP_TO_SINT_SAT, MVT::i64, Custom); 478 479 // Variable arguments. 480 setOperationAction(ISD::VASTART, MVT::Other, Custom); 481 setOperationAction(ISD::VAARG, MVT::Other, Custom); 482 setOperationAction(ISD::VACOPY, MVT::Other, Custom); 483 setOperationAction(ISD::VAEND, MVT::Other, Expand); 484 485 // Variable-sized objects. 486 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); 487 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); 488 489 if (Subtarget->isTargetWindows()) 490 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Custom); 491 else 492 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand); 493 494 // Constant pool entries 495 setOperationAction(ISD::ConstantPool, MVT::i64, Custom); 496 497 // BlockAddress 498 setOperationAction(ISD::BlockAddress, MVT::i64, Custom); 499 500 // Add/Sub overflow ops with MVT::Glues are lowered to NZCV dependences. 501 setOperationAction(ISD::ADDC, MVT::i32, Custom); 502 setOperationAction(ISD::ADDE, MVT::i32, Custom); 503 setOperationAction(ISD::SUBC, MVT::i32, Custom); 504 setOperationAction(ISD::SUBE, MVT::i32, Custom); 505 setOperationAction(ISD::ADDC, MVT::i64, Custom); 506 setOperationAction(ISD::ADDE, MVT::i64, Custom); 507 setOperationAction(ISD::SUBC, MVT::i64, Custom); 508 setOperationAction(ISD::SUBE, MVT::i64, Custom); 509 510 // AArch64 lacks both left-rotate and popcount instructions. 511 setOperationAction(ISD::ROTL, MVT::i32, Expand); 512 setOperationAction(ISD::ROTL, MVT::i64, Expand); 513 for (MVT VT : MVT::fixedlen_vector_valuetypes()) { 514 setOperationAction(ISD::ROTL, VT, Expand); 515 setOperationAction(ISD::ROTR, VT, Expand); 516 } 517 518 // AArch64 doesn't have i32 MULH{S|U}. 519 setOperationAction(ISD::MULHU, MVT::i32, Expand); 520 setOperationAction(ISD::MULHS, MVT::i32, Expand); 521 522 // AArch64 doesn't have {U|S}MUL_LOHI. 523 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand); 524 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand); 525 526 setOperationAction(ISD::CTPOP, MVT::i32, Custom); 527 setOperationAction(ISD::CTPOP, MVT::i64, Custom); 528 setOperationAction(ISD::CTPOP, MVT::i128, Custom); 529 530 setOperationAction(ISD::ABS, MVT::i32, Custom); 531 setOperationAction(ISD::ABS, MVT::i64, Custom); 532 533 setOperationAction(ISD::SDIVREM, MVT::i32, Expand); 534 setOperationAction(ISD::SDIVREM, MVT::i64, Expand); 535 for (MVT VT : MVT::fixedlen_vector_valuetypes()) { 536 setOperationAction(ISD::SDIVREM, VT, Expand); 537 setOperationAction(ISD::UDIVREM, VT, Expand); 538 } 539 setOperationAction(ISD::SREM, MVT::i32, Expand); 540 setOperationAction(ISD::SREM, MVT::i64, Expand); 541 setOperationAction(ISD::UDIVREM, MVT::i32, Expand); 542 setOperationAction(ISD::UDIVREM, MVT::i64, Expand); 543 setOperationAction(ISD::UREM, MVT::i32, Expand); 544 setOperationAction(ISD::UREM, MVT::i64, Expand); 545 546 // Custom lower Add/Sub/Mul with overflow. 547 setOperationAction(ISD::SADDO, MVT::i32, Custom); 548 setOperationAction(ISD::SADDO, MVT::i64, Custom); 549 setOperationAction(ISD::UADDO, MVT::i32, Custom); 550 setOperationAction(ISD::UADDO, MVT::i64, Custom); 551 setOperationAction(ISD::SSUBO, MVT::i32, Custom); 552 setOperationAction(ISD::SSUBO, MVT::i64, Custom); 553 setOperationAction(ISD::USUBO, MVT::i32, Custom); 554 setOperationAction(ISD::USUBO, MVT::i64, Custom); 555 setOperationAction(ISD::SMULO, MVT::i32, Custom); 556 setOperationAction(ISD::SMULO, MVT::i64, Custom); 557 setOperationAction(ISD::UMULO, MVT::i32, Custom); 558 setOperationAction(ISD::UMULO, MVT::i64, Custom); 559 560 setOperationAction(ISD::FSIN, MVT::f32, Expand); 561 setOperationAction(ISD::FSIN, MVT::f64, Expand); 562 setOperationAction(ISD::FCOS, MVT::f32, Expand); 563 setOperationAction(ISD::FCOS, MVT::f64, Expand); 564 setOperationAction(ISD::FPOW, MVT::f32, Expand); 565 setOperationAction(ISD::FPOW, MVT::f64, Expand); 566 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom); 567 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom); 568 if (Subtarget->hasFullFP16()) 569 setOperationAction(ISD::FCOPYSIGN, MVT::f16, Custom); 570 else 571 setOperationAction(ISD::FCOPYSIGN, MVT::f16, Promote); 572 573 setOperationAction(ISD::FREM, MVT::f16, Promote); 574 setOperationAction(ISD::FREM, MVT::v4f16, Expand); 575 setOperationAction(ISD::FREM, MVT::v8f16, Expand); 576 setOperationAction(ISD::FPOW, MVT::f16, Promote); 577 setOperationAction(ISD::FPOW, MVT::v4f16, Expand); 578 setOperationAction(ISD::FPOW, MVT::v8f16, Expand); 579 setOperationAction(ISD::FPOWI, MVT::f16, Promote); 580 setOperationAction(ISD::FPOWI, MVT::v4f16, Expand); 581 setOperationAction(ISD::FPOWI, MVT::v8f16, Expand); 582 setOperationAction(ISD::FCOS, MVT::f16, Promote); 583 setOperationAction(ISD::FCOS, MVT::v4f16, Expand); 584 setOperationAction(ISD::FCOS, MVT::v8f16, Expand); 585 setOperationAction(ISD::FSIN, MVT::f16, Promote); 586 setOperationAction(ISD::FSIN, MVT::v4f16, Expand); 587 setOperationAction(ISD::FSIN, MVT::v8f16, Expand); 588 setOperationAction(ISD::FSINCOS, MVT::f16, Promote); 589 setOperationAction(ISD::FSINCOS, MVT::v4f16, Expand); 590 setOperationAction(ISD::FSINCOS, MVT::v8f16, Expand); 591 setOperationAction(ISD::FEXP, MVT::f16, Promote); 592 setOperationAction(ISD::FEXP, MVT::v4f16, Expand); 593 setOperationAction(ISD::FEXP, MVT::v8f16, Expand); 594 setOperationAction(ISD::FEXP2, MVT::f16, Promote); 595 setOperationAction(ISD::FEXP2, MVT::v4f16, Expand); 596 setOperationAction(ISD::FEXP2, MVT::v8f16, Expand); 597 setOperationAction(ISD::FLOG, MVT::f16, Promote); 598 setOperationAction(ISD::FLOG, MVT::v4f16, Expand); 599 setOperationAction(ISD::FLOG, MVT::v8f16, Expand); 600 setOperationAction(ISD::FLOG2, MVT::f16, Promote); 601 setOperationAction(ISD::FLOG2, MVT::v4f16, Expand); 602 setOperationAction(ISD::FLOG2, MVT::v8f16, Expand); 603 setOperationAction(ISD::FLOG10, MVT::f16, Promote); 604 setOperationAction(ISD::FLOG10, MVT::v4f16, Expand); 605 setOperationAction(ISD::FLOG10, MVT::v8f16, Expand); 606 607 if (!Subtarget->hasFullFP16()) { 608 setOperationAction(ISD::SELECT, MVT::f16, Promote); 609 setOperationAction(ISD::SELECT_CC, MVT::f16, Promote); 610 setOperationAction(ISD::SETCC, MVT::f16, Promote); 611 setOperationAction(ISD::BR_CC, MVT::f16, Promote); 612 setOperationAction(ISD::FADD, MVT::f16, Promote); 613 setOperationAction(ISD::FSUB, MVT::f16, Promote); 614 setOperationAction(ISD::FMUL, MVT::f16, Promote); 615 setOperationAction(ISD::FDIV, MVT::f16, Promote); 616 setOperationAction(ISD::FMA, MVT::f16, Promote); 617 setOperationAction(ISD::FNEG, MVT::f16, Promote); 618 setOperationAction(ISD::FABS, MVT::f16, Promote); 619 setOperationAction(ISD::FCEIL, MVT::f16, Promote); 620 setOperationAction(ISD::FSQRT, MVT::f16, Promote); 621 setOperationAction(ISD::FFLOOR, MVT::f16, Promote); 622 setOperationAction(ISD::FNEARBYINT, MVT::f16, Promote); 623 setOperationAction(ISD::FRINT, MVT::f16, Promote); 624 setOperationAction(ISD::FROUND, MVT::f16, Promote); 625 setOperationAction(ISD::FROUNDEVEN, MVT::f16, Promote); 626 setOperationAction(ISD::FTRUNC, MVT::f16, Promote); 627 setOperationAction(ISD::FMINNUM, MVT::f16, Promote); 628 setOperationAction(ISD::FMAXNUM, MVT::f16, Promote); 629 setOperationAction(ISD::FMINIMUM, MVT::f16, Promote); 630 setOperationAction(ISD::FMAXIMUM, MVT::f16, Promote); 631 632 // promote v4f16 to v4f32 when that is known to be safe. 633 setOperationAction(ISD::FADD, MVT::v4f16, Promote); 634 setOperationAction(ISD::FSUB, MVT::v4f16, Promote); 635 setOperationAction(ISD::FMUL, MVT::v4f16, Promote); 636 setOperationAction(ISD::FDIV, MVT::v4f16, Promote); 637 AddPromotedToType(ISD::FADD, MVT::v4f16, MVT::v4f32); 638 AddPromotedToType(ISD::FSUB, MVT::v4f16, MVT::v4f32); 639 AddPromotedToType(ISD::FMUL, MVT::v4f16, MVT::v4f32); 640 AddPromotedToType(ISD::FDIV, MVT::v4f16, MVT::v4f32); 641 642 setOperationAction(ISD::FABS, MVT::v4f16, Expand); 643 setOperationAction(ISD::FNEG, MVT::v4f16, Expand); 644 setOperationAction(ISD::FROUND, MVT::v4f16, Expand); 645 setOperationAction(ISD::FROUNDEVEN, MVT::v4f16, Expand); 646 setOperationAction(ISD::FMA, MVT::v4f16, Expand); 647 setOperationAction(ISD::SETCC, MVT::v4f16, Expand); 648 setOperationAction(ISD::BR_CC, MVT::v4f16, Expand); 649 setOperationAction(ISD::SELECT, MVT::v4f16, Expand); 650 setOperationAction(ISD::SELECT_CC, MVT::v4f16, Expand); 651 setOperationAction(ISD::FTRUNC, MVT::v4f16, Expand); 652 setOperationAction(ISD::FCOPYSIGN, MVT::v4f16, Expand); 653 setOperationAction(ISD::FFLOOR, MVT::v4f16, Expand); 654 setOperationAction(ISD::FCEIL, MVT::v4f16, Expand); 655 setOperationAction(ISD::FRINT, MVT::v4f16, Expand); 656 setOperationAction(ISD::FNEARBYINT, MVT::v4f16, Expand); 657 setOperationAction(ISD::FSQRT, MVT::v4f16, Expand); 658 659 setOperationAction(ISD::FABS, MVT::v8f16, Expand); 660 setOperationAction(ISD::FADD, MVT::v8f16, Expand); 661 setOperationAction(ISD::FCEIL, MVT::v8f16, Expand); 662 setOperationAction(ISD::FCOPYSIGN, MVT::v8f16, Expand); 663 setOperationAction(ISD::FDIV, MVT::v8f16, Expand); 664 setOperationAction(ISD::FFLOOR, MVT::v8f16, Expand); 665 setOperationAction(ISD::FMA, MVT::v8f16, Expand); 666 setOperationAction(ISD::FMUL, MVT::v8f16, Expand); 667 setOperationAction(ISD::FNEARBYINT, MVT::v8f16, Expand); 668 setOperationAction(ISD::FNEG, MVT::v8f16, Expand); 669 setOperationAction(ISD::FROUND, MVT::v8f16, Expand); 670 setOperationAction(ISD::FROUNDEVEN, MVT::v8f16, Expand); 671 setOperationAction(ISD::FRINT, MVT::v8f16, Expand); 672 setOperationAction(ISD::FSQRT, MVT::v8f16, Expand); 673 setOperationAction(ISD::FSUB, MVT::v8f16, Expand); 674 setOperationAction(ISD::FTRUNC, MVT::v8f16, Expand); 675 setOperationAction(ISD::SETCC, MVT::v8f16, Expand); 676 setOperationAction(ISD::BR_CC, MVT::v8f16, Expand); 677 setOperationAction(ISD::SELECT, MVT::v8f16, Expand); 678 setOperationAction(ISD::SELECT_CC, MVT::v8f16, Expand); 679 setOperationAction(ISD::FP_EXTEND, MVT::v8f16, Expand); 680 } 681 682 // AArch64 has implementations of a lot of rounding-like FP operations. 683 for (MVT Ty : {MVT::f32, MVT::f64}) { 684 setOperationAction(ISD::FFLOOR, Ty, Legal); 685 setOperationAction(ISD::FNEARBYINT, Ty, Legal); 686 setOperationAction(ISD::FCEIL, Ty, Legal); 687 setOperationAction(ISD::FRINT, Ty, Legal); 688 setOperationAction(ISD::FTRUNC, Ty, Legal); 689 setOperationAction(ISD::FROUND, Ty, Legal); 690 setOperationAction(ISD::FROUNDEVEN, Ty, Legal); 691 setOperationAction(ISD::FMINNUM, Ty, Legal); 692 setOperationAction(ISD::FMAXNUM, Ty, Legal); 693 setOperationAction(ISD::FMINIMUM, Ty, Legal); 694 setOperationAction(ISD::FMAXIMUM, Ty, Legal); 695 setOperationAction(ISD::LROUND, Ty, Legal); 696 setOperationAction(ISD::LLROUND, Ty, Legal); 697 setOperationAction(ISD::LRINT, Ty, Legal); 698 setOperationAction(ISD::LLRINT, Ty, Legal); 699 } 700 701 if (Subtarget->hasFullFP16()) { 702 setOperationAction(ISD::FNEARBYINT, MVT::f16, Legal); 703 setOperationAction(ISD::FFLOOR, MVT::f16, Legal); 704 setOperationAction(ISD::FCEIL, MVT::f16, Legal); 705 setOperationAction(ISD::FRINT, MVT::f16, Legal); 706 setOperationAction(ISD::FTRUNC, MVT::f16, Legal); 707 setOperationAction(ISD::FROUND, MVT::f16, Legal); 708 setOperationAction(ISD::FROUNDEVEN, MVT::f16, Legal); 709 setOperationAction(ISD::FMINNUM, MVT::f16, Legal); 710 setOperationAction(ISD::FMAXNUM, MVT::f16, Legal); 711 setOperationAction(ISD::FMINIMUM, MVT::f16, Legal); 712 setOperationAction(ISD::FMAXIMUM, MVT::f16, Legal); 713 } 714 715 setOperationAction(ISD::PREFETCH, MVT::Other, Custom); 716 717 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom); 718 setOperationAction(ISD::SET_ROUNDING, MVT::Other, Custom); 719 720 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i128, Custom); 721 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom); 722 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom); 723 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Custom); 724 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom); 725 726 // Generate outline atomics library calls only if LSE was not specified for 727 // subtarget 728 if (Subtarget->outlineAtomics() && !Subtarget->hasLSE()) { 729 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, LibCall); 730 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, LibCall); 731 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, LibCall); 732 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, LibCall); 733 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i128, LibCall); 734 setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, LibCall); 735 setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, LibCall); 736 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, LibCall); 737 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, LibCall); 738 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, LibCall); 739 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, LibCall); 740 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, LibCall); 741 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, LibCall); 742 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, LibCall); 743 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, LibCall); 744 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, LibCall); 745 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, LibCall); 746 setOperationAction(ISD::ATOMIC_LOAD_CLR, MVT::i8, LibCall); 747 setOperationAction(ISD::ATOMIC_LOAD_CLR, MVT::i16, LibCall); 748 setOperationAction(ISD::ATOMIC_LOAD_CLR, MVT::i32, LibCall); 749 setOperationAction(ISD::ATOMIC_LOAD_CLR, MVT::i64, LibCall); 750 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, LibCall); 751 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, LibCall); 752 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, LibCall); 753 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, LibCall); 754 #define LCALLNAMES(A, B, N) \ 755 setLibcallName(A##N##_RELAX, #B #N "_relax"); \ 756 setLibcallName(A##N##_ACQ, #B #N "_acq"); \ 757 setLibcallName(A##N##_REL, #B #N "_rel"); \ 758 setLibcallName(A##N##_ACQ_REL, #B #N "_acq_rel"); 759 #define LCALLNAME4(A, B) \ 760 LCALLNAMES(A, B, 1) \ 761 LCALLNAMES(A, B, 2) LCALLNAMES(A, B, 4) LCALLNAMES(A, B, 8) 762 #define LCALLNAME5(A, B) \ 763 LCALLNAMES(A, B, 1) \ 764 LCALLNAMES(A, B, 2) \ 765 LCALLNAMES(A, B, 4) LCALLNAMES(A, B, 8) LCALLNAMES(A, B, 16) 766 LCALLNAME5(RTLIB::OUTLINE_ATOMIC_CAS, __aarch64_cas) 767 LCALLNAME4(RTLIB::OUTLINE_ATOMIC_SWP, __aarch64_swp) 768 LCALLNAME4(RTLIB::OUTLINE_ATOMIC_LDADD, __aarch64_ldadd) 769 LCALLNAME4(RTLIB::OUTLINE_ATOMIC_LDSET, __aarch64_ldset) 770 LCALLNAME4(RTLIB::OUTLINE_ATOMIC_LDCLR, __aarch64_ldclr) 771 LCALLNAME4(RTLIB::OUTLINE_ATOMIC_LDEOR, __aarch64_ldeor) 772 #undef LCALLNAMES 773 #undef LCALLNAME4 774 #undef LCALLNAME5 775 } 776 777 // 128-bit loads and stores can be done without expanding 778 setOperationAction(ISD::LOAD, MVT::i128, Custom); 779 setOperationAction(ISD::STORE, MVT::i128, Custom); 780 781 // 256 bit non-temporal stores can be lowered to STNP. Do this as part of the 782 // custom lowering, as there are no un-paired non-temporal stores and 783 // legalization will break up 256 bit inputs. 784 setOperationAction(ISD::STORE, MVT::v32i8, Custom); 785 setOperationAction(ISD::STORE, MVT::v16i16, Custom); 786 setOperationAction(ISD::STORE, MVT::v16f16, Custom); 787 setOperationAction(ISD::STORE, MVT::v8i32, Custom); 788 setOperationAction(ISD::STORE, MVT::v8f32, Custom); 789 setOperationAction(ISD::STORE, MVT::v4f64, Custom); 790 setOperationAction(ISD::STORE, MVT::v4i64, Custom); 791 792 // Lower READCYCLECOUNTER using an mrs from PMCCNTR_EL0. 793 // This requires the Performance Monitors extension. 794 if (Subtarget->hasPerfMon()) 795 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Legal); 796 797 if (getLibcallName(RTLIB::SINCOS_STRET_F32) != nullptr && 798 getLibcallName(RTLIB::SINCOS_STRET_F64) != nullptr) { 799 // Issue __sincos_stret if available. 800 setOperationAction(ISD::FSINCOS, MVT::f64, Custom); 801 setOperationAction(ISD::FSINCOS, MVT::f32, Custom); 802 } else { 803 setOperationAction(ISD::FSINCOS, MVT::f64, Expand); 804 setOperationAction(ISD::FSINCOS, MVT::f32, Expand); 805 } 806 807 if (Subtarget->getTargetTriple().isOSMSVCRT()) { 808 // MSVCRT doesn't have powi; fall back to pow 809 setLibcallName(RTLIB::POWI_F32, nullptr); 810 setLibcallName(RTLIB::POWI_F64, nullptr); 811 } 812 813 // Make floating-point constants legal for the large code model, so they don't 814 // become loads from the constant pool. 815 if (Subtarget->isTargetMachO() && TM.getCodeModel() == CodeModel::Large) { 816 setOperationAction(ISD::ConstantFP, MVT::f32, Legal); 817 setOperationAction(ISD::ConstantFP, MVT::f64, Legal); 818 } 819 820 // AArch64 does not have floating-point extending loads, i1 sign-extending 821 // load, floating-point truncating stores, or v2i32->v2i16 truncating store. 822 for (MVT VT : MVT::fp_valuetypes()) { 823 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand); 824 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand); 825 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f64, Expand); 826 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f80, Expand); 827 } 828 for (MVT VT : MVT::integer_valuetypes()) 829 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Expand); 830 831 setTruncStoreAction(MVT::f32, MVT::f16, Expand); 832 setTruncStoreAction(MVT::f64, MVT::f32, Expand); 833 setTruncStoreAction(MVT::f64, MVT::f16, Expand); 834 setTruncStoreAction(MVT::f128, MVT::f80, Expand); 835 setTruncStoreAction(MVT::f128, MVT::f64, Expand); 836 setTruncStoreAction(MVT::f128, MVT::f32, Expand); 837 setTruncStoreAction(MVT::f128, MVT::f16, Expand); 838 839 setOperationAction(ISD::BITCAST, MVT::i16, Custom); 840 setOperationAction(ISD::BITCAST, MVT::f16, Custom); 841 setOperationAction(ISD::BITCAST, MVT::bf16, Custom); 842 843 // Indexed loads and stores are supported. 844 for (unsigned im = (unsigned)ISD::PRE_INC; 845 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) { 846 setIndexedLoadAction(im, MVT::i8, Legal); 847 setIndexedLoadAction(im, MVT::i16, Legal); 848 setIndexedLoadAction(im, MVT::i32, Legal); 849 setIndexedLoadAction(im, MVT::i64, Legal); 850 setIndexedLoadAction(im, MVT::f64, Legal); 851 setIndexedLoadAction(im, MVT::f32, Legal); 852 setIndexedLoadAction(im, MVT::f16, Legal); 853 setIndexedLoadAction(im, MVT::bf16, Legal); 854 setIndexedStoreAction(im, MVT::i8, Legal); 855 setIndexedStoreAction(im, MVT::i16, Legal); 856 setIndexedStoreAction(im, MVT::i32, Legal); 857 setIndexedStoreAction(im, MVT::i64, Legal); 858 setIndexedStoreAction(im, MVT::f64, Legal); 859 setIndexedStoreAction(im, MVT::f32, Legal); 860 setIndexedStoreAction(im, MVT::f16, Legal); 861 setIndexedStoreAction(im, MVT::bf16, Legal); 862 } 863 864 // Trap. 865 setOperationAction(ISD::TRAP, MVT::Other, Legal); 866 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal); 867 setOperationAction(ISD::UBSANTRAP, MVT::Other, Legal); 868 869 // We combine OR nodes for bitfield operations. 870 setTargetDAGCombine(ISD::OR); 871 // Try to create BICs for vector ANDs. 872 setTargetDAGCombine(ISD::AND); 873 874 // Vector add and sub nodes may conceal a high-half opportunity. 875 // Also, try to fold ADD into CSINC/CSINV.. 876 setTargetDAGCombine(ISD::ADD); 877 setTargetDAGCombine(ISD::ABS); 878 setTargetDAGCombine(ISD::SUB); 879 setTargetDAGCombine(ISD::SRL); 880 setTargetDAGCombine(ISD::XOR); 881 setTargetDAGCombine(ISD::SINT_TO_FP); 882 setTargetDAGCombine(ISD::UINT_TO_FP); 883 884 // TODO: Do the same for FP_TO_*INT_SAT. 885 setTargetDAGCombine(ISD::FP_TO_SINT); 886 setTargetDAGCombine(ISD::FP_TO_UINT); 887 setTargetDAGCombine(ISD::FDIV); 888 889 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN); 890 891 setTargetDAGCombine(ISD::ANY_EXTEND); 892 setTargetDAGCombine(ISD::ZERO_EXTEND); 893 setTargetDAGCombine(ISD::SIGN_EXTEND); 894 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG); 895 setTargetDAGCombine(ISD::TRUNCATE); 896 setTargetDAGCombine(ISD::CONCAT_VECTORS); 897 setTargetDAGCombine(ISD::STORE); 898 if (Subtarget->supportsAddressTopByteIgnored()) 899 setTargetDAGCombine(ISD::LOAD); 900 901 setTargetDAGCombine(ISD::MUL); 902 903 setTargetDAGCombine(ISD::SELECT); 904 setTargetDAGCombine(ISD::VSELECT); 905 906 setTargetDAGCombine(ISD::INTRINSIC_VOID); 907 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN); 908 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT); 909 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT); 910 setTargetDAGCombine(ISD::VECREDUCE_ADD); 911 setTargetDAGCombine(ISD::STEP_VECTOR); 912 913 setTargetDAGCombine(ISD::GlobalAddress); 914 915 // In case of strict alignment, avoid an excessive number of byte wide stores. 916 MaxStoresPerMemsetOptSize = 8; 917 MaxStoresPerMemset = Subtarget->requiresStrictAlign() 918 ? MaxStoresPerMemsetOptSize : 32; 919 920 MaxGluedStoresPerMemcpy = 4; 921 MaxStoresPerMemcpyOptSize = 4; 922 MaxStoresPerMemcpy = Subtarget->requiresStrictAlign() 923 ? MaxStoresPerMemcpyOptSize : 16; 924 925 MaxStoresPerMemmoveOptSize = MaxStoresPerMemmove = 4; 926 927 MaxLoadsPerMemcmpOptSize = 4; 928 MaxLoadsPerMemcmp = Subtarget->requiresStrictAlign() 929 ? MaxLoadsPerMemcmpOptSize : 8; 930 931 setStackPointerRegisterToSaveRestore(AArch64::SP); 932 933 setSchedulingPreference(Sched::Hybrid); 934 935 EnableExtLdPromotion = true; 936 937 // Set required alignment. 938 setMinFunctionAlignment(Align(4)); 939 // Set preferred alignments. 940 setPrefLoopAlignment(Align(1ULL << STI.getPrefLoopLogAlignment())); 941 setPrefFunctionAlignment(Align(1ULL << STI.getPrefFunctionLogAlignment())); 942 943 // Only change the limit for entries in a jump table if specified by 944 // the sub target, but not at the command line. 945 unsigned MaxJT = STI.getMaximumJumpTableSize(); 946 if (MaxJT && getMaximumJumpTableSize() == UINT_MAX) 947 setMaximumJumpTableSize(MaxJT); 948 949 setHasExtractBitsInsn(true); 950 951 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom); 952 953 if (Subtarget->hasNEON()) { 954 // FIXME: v1f64 shouldn't be legal if we can avoid it, because it leads to 955 // silliness like this: 956 setOperationAction(ISD::FABS, MVT::v1f64, Expand); 957 setOperationAction(ISD::FADD, MVT::v1f64, Expand); 958 setOperationAction(ISD::FCEIL, MVT::v1f64, Expand); 959 setOperationAction(ISD::FCOPYSIGN, MVT::v1f64, Expand); 960 setOperationAction(ISD::FCOS, MVT::v1f64, Expand); 961 setOperationAction(ISD::FDIV, MVT::v1f64, Expand); 962 setOperationAction(ISD::FFLOOR, MVT::v1f64, Expand); 963 setOperationAction(ISD::FMA, MVT::v1f64, Expand); 964 setOperationAction(ISD::FMUL, MVT::v1f64, Expand); 965 setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Expand); 966 setOperationAction(ISD::FNEG, MVT::v1f64, Expand); 967 setOperationAction(ISD::FPOW, MVT::v1f64, Expand); 968 setOperationAction(ISD::FREM, MVT::v1f64, Expand); 969 setOperationAction(ISD::FROUND, MVT::v1f64, Expand); 970 setOperationAction(ISD::FROUNDEVEN, MVT::v1f64, Expand); 971 setOperationAction(ISD::FRINT, MVT::v1f64, Expand); 972 setOperationAction(ISD::FSIN, MVT::v1f64, Expand); 973 setOperationAction(ISD::FSINCOS, MVT::v1f64, Expand); 974 setOperationAction(ISD::FSQRT, MVT::v1f64, Expand); 975 setOperationAction(ISD::FSUB, MVT::v1f64, Expand); 976 setOperationAction(ISD::FTRUNC, MVT::v1f64, Expand); 977 setOperationAction(ISD::SETCC, MVT::v1f64, Expand); 978 setOperationAction(ISD::BR_CC, MVT::v1f64, Expand); 979 setOperationAction(ISD::SELECT, MVT::v1f64, Expand); 980 setOperationAction(ISD::SELECT_CC, MVT::v1f64, Expand); 981 setOperationAction(ISD::FP_EXTEND, MVT::v1f64, Expand); 982 983 setOperationAction(ISD::FP_TO_SINT, MVT::v1i64, Expand); 984 setOperationAction(ISD::FP_TO_UINT, MVT::v1i64, Expand); 985 setOperationAction(ISD::SINT_TO_FP, MVT::v1i64, Expand); 986 setOperationAction(ISD::UINT_TO_FP, MVT::v1i64, Expand); 987 setOperationAction(ISD::FP_ROUND, MVT::v1f64, Expand); 988 989 setOperationAction(ISD::MUL, MVT::v1i64, Expand); 990 991 // AArch64 doesn't have a direct vector ->f32 conversion instructions for 992 // elements smaller than i32, so promote the input to i32 first. 993 setOperationPromotedToType(ISD::UINT_TO_FP, MVT::v4i8, MVT::v4i32); 994 setOperationPromotedToType(ISD::SINT_TO_FP, MVT::v4i8, MVT::v4i32); 995 setOperationPromotedToType(ISD::SINT_TO_FP, MVT::v8i8, MVT::v8i32); 996 setOperationPromotedToType(ISD::UINT_TO_FP, MVT::v8i8, MVT::v8i32); 997 setOperationPromotedToType(ISD::UINT_TO_FP, MVT::v16i8, MVT::v16i32); 998 setOperationPromotedToType(ISD::SINT_TO_FP, MVT::v16i8, MVT::v16i32); 999 1000 // Similarly, there is no direct i32 -> f64 vector conversion instruction. 1001 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom); 1002 setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom); 1003 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom); 1004 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom); 1005 // Or, direct i32 -> f16 vector conversion. Set it so custom, so the 1006 // conversion happens in two steps: v4i32 -> v4f32 -> v4f16 1007 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Custom); 1008 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom); 1009 1010 if (Subtarget->hasFullFP16()) { 1011 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom); 1012 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom); 1013 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Custom); 1014 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom); 1015 } else { 1016 // when AArch64 doesn't have fullfp16 support, promote the input 1017 // to i32 first. 1018 setOperationPromotedToType(ISD::UINT_TO_FP, MVT::v4i16, MVT::v4i32); 1019 setOperationPromotedToType(ISD::SINT_TO_FP, MVT::v4i16, MVT::v4i32); 1020 setOperationPromotedToType(ISD::SINT_TO_FP, MVT::v8i16, MVT::v8i32); 1021 setOperationPromotedToType(ISD::UINT_TO_FP, MVT::v8i16, MVT::v8i32); 1022 } 1023 1024 setOperationAction(ISD::CTLZ, MVT::v1i64, Expand); 1025 setOperationAction(ISD::CTLZ, MVT::v2i64, Expand); 1026 setOperationAction(ISD::BITREVERSE, MVT::v8i8, Legal); 1027 setOperationAction(ISD::BITREVERSE, MVT::v16i8, Legal); 1028 1029 // AArch64 doesn't have MUL.2d: 1030 setOperationAction(ISD::MUL, MVT::v2i64, Expand); 1031 // Custom handling for some quad-vector types to detect MULL. 1032 setOperationAction(ISD::MUL, MVT::v8i16, Custom); 1033 setOperationAction(ISD::MUL, MVT::v4i32, Custom); 1034 setOperationAction(ISD::MUL, MVT::v2i64, Custom); 1035 1036 // Saturates 1037 for (MVT VT : { MVT::v8i8, MVT::v4i16, MVT::v2i32, 1038 MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64 }) { 1039 setOperationAction(ISD::SADDSAT, VT, Legal); 1040 setOperationAction(ISD::UADDSAT, VT, Legal); 1041 setOperationAction(ISD::SSUBSAT, VT, Legal); 1042 setOperationAction(ISD::USUBSAT, VT, Legal); 1043 } 1044 1045 // Vector reductions 1046 for (MVT VT : { MVT::v4f16, MVT::v2f32, 1047 MVT::v8f16, MVT::v4f32, MVT::v2f64 }) { 1048 if (VT.getVectorElementType() != MVT::f16 || Subtarget->hasFullFP16()) { 1049 setOperationAction(ISD::VECREDUCE_FMAX, VT, Custom); 1050 setOperationAction(ISD::VECREDUCE_FMIN, VT, Custom); 1051 1052 setOperationAction(ISD::VECREDUCE_FADD, VT, Legal); 1053 } 1054 } 1055 for (MVT VT : { MVT::v8i8, MVT::v4i16, MVT::v2i32, 1056 MVT::v16i8, MVT::v8i16, MVT::v4i32 }) { 1057 setOperationAction(ISD::VECREDUCE_ADD, VT, Custom); 1058 setOperationAction(ISD::VECREDUCE_SMAX, VT, Custom); 1059 setOperationAction(ISD::VECREDUCE_SMIN, VT, Custom); 1060 setOperationAction(ISD::VECREDUCE_UMAX, VT, Custom); 1061 setOperationAction(ISD::VECREDUCE_UMIN, VT, Custom); 1062 } 1063 setOperationAction(ISD::VECREDUCE_ADD, MVT::v2i64, Custom); 1064 1065 setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Legal); 1066 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand); 1067 // Likewise, narrowing and extending vector loads/stores aren't handled 1068 // directly. 1069 for (MVT VT : MVT::fixedlen_vector_valuetypes()) { 1070 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand); 1071 1072 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32) { 1073 setOperationAction(ISD::MULHS, VT, Legal); 1074 setOperationAction(ISD::MULHU, VT, Legal); 1075 } else { 1076 setOperationAction(ISD::MULHS, VT, Expand); 1077 setOperationAction(ISD::MULHU, VT, Expand); 1078 } 1079 setOperationAction(ISD::SMUL_LOHI, VT, Expand); 1080 setOperationAction(ISD::UMUL_LOHI, VT, Expand); 1081 1082 setOperationAction(ISD::BSWAP, VT, Expand); 1083 setOperationAction(ISD::CTTZ, VT, Expand); 1084 1085 for (MVT InnerVT : MVT::fixedlen_vector_valuetypes()) { 1086 setTruncStoreAction(VT, InnerVT, Expand); 1087 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand); 1088 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand); 1089 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand); 1090 } 1091 } 1092 1093 // AArch64 has implementations of a lot of rounding-like FP operations. 1094 for (MVT Ty : {MVT::v2f32, MVT::v4f32, MVT::v2f64}) { 1095 setOperationAction(ISD::FFLOOR, Ty, Legal); 1096 setOperationAction(ISD::FNEARBYINT, Ty, Legal); 1097 setOperationAction(ISD::FCEIL, Ty, Legal); 1098 setOperationAction(ISD::FRINT, Ty, Legal); 1099 setOperationAction(ISD::FTRUNC, Ty, Legal); 1100 setOperationAction(ISD::FROUND, Ty, Legal); 1101 setOperationAction(ISD::FROUNDEVEN, Ty, Legal); 1102 } 1103 1104 if (Subtarget->hasFullFP16()) { 1105 for (MVT Ty : {MVT::v4f16, MVT::v8f16}) { 1106 setOperationAction(ISD::FFLOOR, Ty, Legal); 1107 setOperationAction(ISD::FNEARBYINT, Ty, Legal); 1108 setOperationAction(ISD::FCEIL, Ty, Legal); 1109 setOperationAction(ISD::FRINT, Ty, Legal); 1110 setOperationAction(ISD::FTRUNC, Ty, Legal); 1111 setOperationAction(ISD::FROUND, Ty, Legal); 1112 setOperationAction(ISD::FROUNDEVEN, Ty, Legal); 1113 } 1114 } 1115 1116 if (Subtarget->hasSVE()) 1117 setOperationAction(ISD::VSCALE, MVT::i32, Custom); 1118 1119 setTruncStoreAction(MVT::v4i16, MVT::v4i8, Custom); 1120 } 1121 1122 if (Subtarget->hasSVE()) { 1123 for (auto VT : {MVT::nxv16i8, MVT::nxv8i16, MVT::nxv4i32, MVT::nxv2i64}) { 1124 setOperationAction(ISD::BITREVERSE, VT, Custom); 1125 setOperationAction(ISD::BSWAP, VT, Custom); 1126 setOperationAction(ISD::CTLZ, VT, Custom); 1127 setOperationAction(ISD::CTPOP, VT, Custom); 1128 setOperationAction(ISD::CTTZ, VT, Custom); 1129 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom); 1130 setOperationAction(ISD::UINT_TO_FP, VT, Custom); 1131 setOperationAction(ISD::SINT_TO_FP, VT, Custom); 1132 setOperationAction(ISD::FP_TO_UINT, VT, Custom); 1133 setOperationAction(ISD::FP_TO_SINT, VT, Custom); 1134 setOperationAction(ISD::MGATHER, VT, Custom); 1135 setOperationAction(ISD::MSCATTER, VT, Custom); 1136 setOperationAction(ISD::MUL, VT, Custom); 1137 setOperationAction(ISD::MULHS, VT, Custom); 1138 setOperationAction(ISD::MULHU, VT, Custom); 1139 setOperationAction(ISD::SPLAT_VECTOR, VT, Custom); 1140 setOperationAction(ISD::SELECT, VT, Custom); 1141 setOperationAction(ISD::SETCC, VT, Custom); 1142 setOperationAction(ISD::SDIV, VT, Custom); 1143 setOperationAction(ISD::UDIV, VT, Custom); 1144 setOperationAction(ISD::SMIN, VT, Custom); 1145 setOperationAction(ISD::UMIN, VT, Custom); 1146 setOperationAction(ISD::SMAX, VT, Custom); 1147 setOperationAction(ISD::UMAX, VT, Custom); 1148 setOperationAction(ISD::SHL, VT, Custom); 1149 setOperationAction(ISD::SRL, VT, Custom); 1150 setOperationAction(ISD::SRA, VT, Custom); 1151 setOperationAction(ISD::ABS, VT, Custom); 1152 setOperationAction(ISD::VECREDUCE_ADD, VT, Custom); 1153 setOperationAction(ISD::VECREDUCE_AND, VT, Custom); 1154 setOperationAction(ISD::VECREDUCE_OR, VT, Custom); 1155 setOperationAction(ISD::VECREDUCE_XOR, VT, Custom); 1156 setOperationAction(ISD::VECREDUCE_UMIN, VT, Custom); 1157 setOperationAction(ISD::VECREDUCE_UMAX, VT, Custom); 1158 setOperationAction(ISD::VECREDUCE_SMIN, VT, Custom); 1159 setOperationAction(ISD::VECREDUCE_SMAX, VT, Custom); 1160 1161 setOperationAction(ISD::UMUL_LOHI, VT, Expand); 1162 setOperationAction(ISD::SMUL_LOHI, VT, Expand); 1163 setOperationAction(ISD::SELECT_CC, VT, Expand); 1164 setOperationAction(ISD::ROTL, VT, Expand); 1165 setOperationAction(ISD::ROTR, VT, Expand); 1166 } 1167 1168 // Illegal unpacked integer vector types. 1169 for (auto VT : {MVT::nxv8i8, MVT::nxv4i16, MVT::nxv2i32}) { 1170 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom); 1171 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom); 1172 } 1173 1174 for (auto VT : {MVT::nxv16i1, MVT::nxv8i1, MVT::nxv4i1, MVT::nxv2i1}) { 1175 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom); 1176 setOperationAction(ISD::SELECT, VT, Custom); 1177 setOperationAction(ISD::SETCC, VT, Custom); 1178 setOperationAction(ISD::SPLAT_VECTOR, VT, Custom); 1179 setOperationAction(ISD::TRUNCATE, VT, Custom); 1180 setOperationAction(ISD::VECREDUCE_AND, VT, Custom); 1181 setOperationAction(ISD::VECREDUCE_OR, VT, Custom); 1182 setOperationAction(ISD::VECREDUCE_XOR, VT, Custom); 1183 1184 setOperationAction(ISD::SELECT_CC, VT, Expand); 1185 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); 1186 1187 // There are no legal MVT::nxv16f## based types. 1188 if (VT != MVT::nxv16i1) { 1189 setOperationAction(ISD::SINT_TO_FP, VT, Custom); 1190 setOperationAction(ISD::UINT_TO_FP, VT, Custom); 1191 } 1192 1193 // NEON doesn't support masked loads or stores, but SVE does 1194 for (auto VT : 1195 {MVT::v4f16, MVT::v8f16, MVT::v2f32, MVT::v4f32, MVT::v1f64, 1196 MVT::v2f64, MVT::v8i8, MVT::v16i8, MVT::v4i16, MVT::v8i16, 1197 MVT::v2i32, MVT::v4i32, MVT::v1i64, MVT::v2i64}) { 1198 setOperationAction(ISD::MLOAD, VT, Custom); 1199 setOperationAction(ISD::MSTORE, VT, Custom); 1200 } 1201 } 1202 1203 for (MVT VT : MVT::fp_scalable_vector_valuetypes()) { 1204 for (MVT InnerVT : MVT::fp_scalable_vector_valuetypes()) { 1205 // Avoid marking truncating FP stores as legal to prevent the 1206 // DAGCombiner from creating unsupported truncating stores. 1207 setTruncStoreAction(VT, InnerVT, Expand); 1208 // SVE does not have floating-point extending loads. 1209 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand); 1210 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand); 1211 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand); 1212 } 1213 } 1214 1215 for (auto VT : {MVT::nxv2f16, MVT::nxv4f16, MVT::nxv8f16, MVT::nxv2f32, 1216 MVT::nxv4f32, MVT::nxv2f64}) { 1217 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom); 1218 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom); 1219 setOperationAction(ISD::MGATHER, VT, Custom); 1220 setOperationAction(ISD::MSCATTER, VT, Custom); 1221 setOperationAction(ISD::SPLAT_VECTOR, VT, Custom); 1222 setOperationAction(ISD::SELECT, VT, Custom); 1223 setOperationAction(ISD::FADD, VT, Custom); 1224 setOperationAction(ISD::FDIV, VT, Custom); 1225 setOperationAction(ISD::FMA, VT, Custom); 1226 setOperationAction(ISD::FMAXIMUM, VT, Custom); 1227 setOperationAction(ISD::FMAXNUM, VT, Custom); 1228 setOperationAction(ISD::FMINIMUM, VT, Custom); 1229 setOperationAction(ISD::FMINNUM, VT, Custom); 1230 setOperationAction(ISD::FMUL, VT, Custom); 1231 setOperationAction(ISD::FNEG, VT, Custom); 1232 setOperationAction(ISD::FSUB, VT, Custom); 1233 setOperationAction(ISD::FCEIL, VT, Custom); 1234 setOperationAction(ISD::FFLOOR, VT, Custom); 1235 setOperationAction(ISD::FNEARBYINT, VT, Custom); 1236 setOperationAction(ISD::FRINT, VT, Custom); 1237 setOperationAction(ISD::FROUND, VT, Custom); 1238 setOperationAction(ISD::FROUNDEVEN, VT, Custom); 1239 setOperationAction(ISD::FTRUNC, VT, Custom); 1240 setOperationAction(ISD::FSQRT, VT, Custom); 1241 setOperationAction(ISD::FABS, VT, Custom); 1242 setOperationAction(ISD::FP_EXTEND, VT, Custom); 1243 setOperationAction(ISD::FP_ROUND, VT, Custom); 1244 setOperationAction(ISD::VECREDUCE_FADD, VT, Custom); 1245 setOperationAction(ISD::VECREDUCE_FMAX, VT, Custom); 1246 setOperationAction(ISD::VECREDUCE_FMIN, VT, Custom); 1247 setOperationAction(ISD::VECREDUCE_SEQ_FADD, VT, Custom); 1248 1249 setOperationAction(ISD::SELECT_CC, VT, Expand); 1250 } 1251 1252 for (auto VT : {MVT::nxv2bf16, MVT::nxv4bf16, MVT::nxv8bf16}) { 1253 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom); 1254 setOperationAction(ISD::MGATHER, VT, Custom); 1255 setOperationAction(ISD::MSCATTER, VT, Custom); 1256 } 1257 1258 setOperationAction(ISD::SPLAT_VECTOR, MVT::nxv8bf16, Custom); 1259 1260 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i8, Custom); 1261 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i16, Custom); 1262 1263 // NOTE: Currently this has to happen after computeRegisterProperties rather 1264 // than the preferred option of combining it with the addRegisterClass call. 1265 if (Subtarget->useSVEForFixedLengthVectors()) { 1266 for (MVT VT : MVT::integer_fixedlen_vector_valuetypes()) 1267 if (useSVEForFixedLengthVectorVT(VT)) 1268 addTypeForFixedLengthSVE(VT); 1269 for (MVT VT : MVT::fp_fixedlen_vector_valuetypes()) 1270 if (useSVEForFixedLengthVectorVT(VT)) 1271 addTypeForFixedLengthSVE(VT); 1272 1273 // 64bit results can mean a bigger than NEON input. 1274 for (auto VT : {MVT::v8i8, MVT::v4i16}) 1275 setOperationAction(ISD::TRUNCATE, VT, Custom); 1276 setOperationAction(ISD::FP_ROUND, MVT::v4f16, Custom); 1277 1278 // 128bit results imply a bigger than NEON input. 1279 for (auto VT : {MVT::v16i8, MVT::v8i16, MVT::v4i32}) 1280 setOperationAction(ISD::TRUNCATE, VT, Custom); 1281 for (auto VT : {MVT::v8f16, MVT::v4f32}) 1282 setOperationAction(ISD::FP_ROUND, VT, Expand); 1283 1284 // These operations are not supported on NEON but SVE can do them. 1285 setOperationAction(ISD::BITREVERSE, MVT::v1i64, Custom); 1286 setOperationAction(ISD::CTLZ, MVT::v1i64, Custom); 1287 setOperationAction(ISD::CTLZ, MVT::v2i64, Custom); 1288 setOperationAction(ISD::CTTZ, MVT::v1i64, Custom); 1289 setOperationAction(ISD::MUL, MVT::v1i64, Custom); 1290 setOperationAction(ISD::MUL, MVT::v2i64, Custom); 1291 setOperationAction(ISD::MULHS, MVT::v1i64, Custom); 1292 setOperationAction(ISD::MULHS, MVT::v2i64, Custom); 1293 setOperationAction(ISD::MULHU, MVT::v1i64, Custom); 1294 setOperationAction(ISD::MULHU, MVT::v2i64, Custom); 1295 setOperationAction(ISD::SDIV, MVT::v8i8, Custom); 1296 setOperationAction(ISD::SDIV, MVT::v16i8, Custom); 1297 setOperationAction(ISD::SDIV, MVT::v4i16, Custom); 1298 setOperationAction(ISD::SDIV, MVT::v8i16, Custom); 1299 setOperationAction(ISD::SDIV, MVT::v2i32, Custom); 1300 setOperationAction(ISD::SDIV, MVT::v4i32, Custom); 1301 setOperationAction(ISD::SDIV, MVT::v1i64, Custom); 1302 setOperationAction(ISD::SDIV, MVT::v2i64, Custom); 1303 setOperationAction(ISD::SMAX, MVT::v1i64, Custom); 1304 setOperationAction(ISD::SMAX, MVT::v2i64, Custom); 1305 setOperationAction(ISD::SMIN, MVT::v1i64, Custom); 1306 setOperationAction(ISD::SMIN, MVT::v2i64, Custom); 1307 setOperationAction(ISD::UDIV, MVT::v8i8, Custom); 1308 setOperationAction(ISD::UDIV, MVT::v16i8, Custom); 1309 setOperationAction(ISD::UDIV, MVT::v4i16, Custom); 1310 setOperationAction(ISD::UDIV, MVT::v8i16, Custom); 1311 setOperationAction(ISD::UDIV, MVT::v2i32, Custom); 1312 setOperationAction(ISD::UDIV, MVT::v4i32, Custom); 1313 setOperationAction(ISD::UDIV, MVT::v1i64, Custom); 1314 setOperationAction(ISD::UDIV, MVT::v2i64, Custom); 1315 setOperationAction(ISD::UMAX, MVT::v1i64, Custom); 1316 setOperationAction(ISD::UMAX, MVT::v2i64, Custom); 1317 setOperationAction(ISD::UMIN, MVT::v1i64, Custom); 1318 setOperationAction(ISD::UMIN, MVT::v2i64, Custom); 1319 setOperationAction(ISD::VECREDUCE_SMAX, MVT::v2i64, Custom); 1320 setOperationAction(ISD::VECREDUCE_SMIN, MVT::v2i64, Custom); 1321 setOperationAction(ISD::VECREDUCE_UMAX, MVT::v2i64, Custom); 1322 setOperationAction(ISD::VECREDUCE_UMIN, MVT::v2i64, Custom); 1323 1324 // Int operations with no NEON support. 1325 for (auto VT : {MVT::v8i8, MVT::v16i8, MVT::v4i16, MVT::v8i16, 1326 MVT::v2i32, MVT::v4i32, MVT::v2i64}) { 1327 setOperationAction(ISD::BITREVERSE, VT, Custom); 1328 setOperationAction(ISD::CTTZ, VT, Custom); 1329 setOperationAction(ISD::VECREDUCE_AND, VT, Custom); 1330 setOperationAction(ISD::VECREDUCE_OR, VT, Custom); 1331 setOperationAction(ISD::VECREDUCE_XOR, VT, Custom); 1332 } 1333 1334 // FP operations with no NEON support. 1335 for (auto VT : {MVT::v4f16, MVT::v8f16, MVT::v2f32, MVT::v4f32, 1336 MVT::v1f64, MVT::v2f64}) 1337 setOperationAction(ISD::VECREDUCE_SEQ_FADD, VT, Custom); 1338 1339 // Use SVE for vectors with more than 2 elements. 1340 for (auto VT : {MVT::v4f16, MVT::v8f16, MVT::v4f32}) 1341 setOperationAction(ISD::VECREDUCE_FADD, VT, Custom); 1342 } 1343 1344 setOperationPromotedToType(ISD::VECTOR_SPLICE, MVT::nxv2i1, MVT::nxv2i64); 1345 setOperationPromotedToType(ISD::VECTOR_SPLICE, MVT::nxv4i1, MVT::nxv4i32); 1346 setOperationPromotedToType(ISD::VECTOR_SPLICE, MVT::nxv8i1, MVT::nxv8i16); 1347 setOperationPromotedToType(ISD::VECTOR_SPLICE, MVT::nxv16i1, MVT::nxv16i8); 1348 } 1349 1350 PredictableSelectIsExpensive = Subtarget->predictableSelectIsExpensive(); 1351 } 1352 1353 void AArch64TargetLowering::addTypeForNEON(MVT VT, MVT PromotedBitwiseVT) { 1354 assert(VT.isVector() && "VT should be a vector type"); 1355 1356 if (VT.isFloatingPoint()) { 1357 MVT PromoteTo = EVT(VT).changeVectorElementTypeToInteger().getSimpleVT(); 1358 setOperationPromotedToType(ISD::LOAD, VT, PromoteTo); 1359 setOperationPromotedToType(ISD::STORE, VT, PromoteTo); 1360 } 1361 1362 // Mark vector float intrinsics as expand. 1363 if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64) { 1364 setOperationAction(ISD::FSIN, VT, Expand); 1365 setOperationAction(ISD::FCOS, VT, Expand); 1366 setOperationAction(ISD::FPOW, VT, Expand); 1367 setOperationAction(ISD::FLOG, VT, Expand); 1368 setOperationAction(ISD::FLOG2, VT, Expand); 1369 setOperationAction(ISD::FLOG10, VT, Expand); 1370 setOperationAction(ISD::FEXP, VT, Expand); 1371 setOperationAction(ISD::FEXP2, VT, Expand); 1372 1373 // But we do support custom-lowering for FCOPYSIGN. 1374 setOperationAction(ISD::FCOPYSIGN, VT, Custom); 1375 } 1376 1377 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); 1378 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom); 1379 setOperationAction(ISD::BUILD_VECTOR, VT, Custom); 1380 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom); 1381 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom); 1382 setOperationAction(ISD::SRA, VT, Custom); 1383 setOperationAction(ISD::SRL, VT, Custom); 1384 setOperationAction(ISD::SHL, VT, Custom); 1385 setOperationAction(ISD::OR, VT, Custom); 1386 setOperationAction(ISD::SETCC, VT, Custom); 1387 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal); 1388 1389 setOperationAction(ISD::SELECT, VT, Expand); 1390 setOperationAction(ISD::SELECT_CC, VT, Expand); 1391 setOperationAction(ISD::VSELECT, VT, Expand); 1392 for (MVT InnerVT : MVT::all_valuetypes()) 1393 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand); 1394 1395 // CNT supports only B element sizes, then use UADDLP to widen. 1396 if (VT != MVT::v8i8 && VT != MVT::v16i8) 1397 setOperationAction(ISD::CTPOP, VT, Custom); 1398 1399 setOperationAction(ISD::UDIV, VT, Expand); 1400 setOperationAction(ISD::SDIV, VT, Expand); 1401 setOperationAction(ISD::UREM, VT, Expand); 1402 setOperationAction(ISD::SREM, VT, Expand); 1403 setOperationAction(ISD::FREM, VT, Expand); 1404 1405 setOperationAction(ISD::FP_TO_SINT, VT, Custom); 1406 setOperationAction(ISD::FP_TO_UINT, VT, Custom); 1407 1408 if (!VT.isFloatingPoint()) 1409 setOperationAction(ISD::ABS, VT, Legal); 1410 1411 // [SU][MIN|MAX] are available for all NEON types apart from i64. 1412 if (!VT.isFloatingPoint() && VT != MVT::v2i64 && VT != MVT::v1i64) 1413 for (unsigned Opcode : {ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX}) 1414 setOperationAction(Opcode, VT, Legal); 1415 1416 // F[MIN|MAX][NUM|NAN] are available for all FP NEON types. 1417 if (VT.isFloatingPoint() && 1418 VT.getVectorElementType() != MVT::bf16 && 1419 (VT.getVectorElementType() != MVT::f16 || Subtarget->hasFullFP16())) 1420 for (unsigned Opcode : 1421 {ISD::FMINIMUM, ISD::FMAXIMUM, ISD::FMINNUM, ISD::FMAXNUM}) 1422 setOperationAction(Opcode, VT, Legal); 1423 1424 if (Subtarget->isLittleEndian()) { 1425 for (unsigned im = (unsigned)ISD::PRE_INC; 1426 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) { 1427 setIndexedLoadAction(im, VT, Legal); 1428 setIndexedStoreAction(im, VT, Legal); 1429 } 1430 } 1431 } 1432 1433 void AArch64TargetLowering::addTypeForFixedLengthSVE(MVT VT) { 1434 assert(VT.isFixedLengthVector() && "Expected fixed length vector type!"); 1435 1436 // By default everything must be expanded. 1437 for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op) 1438 setOperationAction(Op, VT, Expand); 1439 1440 // We use EXTRACT_SUBVECTOR to "cast" a scalable vector to a fixed length one. 1441 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom); 1442 1443 if (VT.isFloatingPoint()) { 1444 setCondCodeAction(ISD::SETO, VT, Expand); 1445 setCondCodeAction(ISD::SETOLT, VT, Expand); 1446 setCondCodeAction(ISD::SETLT, VT, Expand); 1447 setCondCodeAction(ISD::SETOLE, VT, Expand); 1448 setCondCodeAction(ISD::SETLE, VT, Expand); 1449 setCondCodeAction(ISD::SETULT, VT, Expand); 1450 setCondCodeAction(ISD::SETULE, VT, Expand); 1451 setCondCodeAction(ISD::SETUGE, VT, Expand); 1452 setCondCodeAction(ISD::SETUGT, VT, Expand); 1453 setCondCodeAction(ISD::SETUEQ, VT, Expand); 1454 setCondCodeAction(ISD::SETUNE, VT, Expand); 1455 } 1456 1457 // Lower fixed length vector operations to scalable equivalents. 1458 setOperationAction(ISD::ABS, VT, Custom); 1459 setOperationAction(ISD::ADD, VT, Custom); 1460 setOperationAction(ISD::AND, VT, Custom); 1461 setOperationAction(ISD::ANY_EXTEND, VT, Custom); 1462 setOperationAction(ISD::BITCAST, VT, Custom); 1463 setOperationAction(ISD::BITREVERSE, VT, Custom); 1464 setOperationAction(ISD::BSWAP, VT, Custom); 1465 setOperationAction(ISD::CTLZ, VT, Custom); 1466 setOperationAction(ISD::CTPOP, VT, Custom); 1467 setOperationAction(ISD::CTTZ, VT, Custom); 1468 setOperationAction(ISD::FABS, VT, Custom); 1469 setOperationAction(ISD::FADD, VT, Custom); 1470 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); 1471 setOperationAction(ISD::FCEIL, VT, Custom); 1472 setOperationAction(ISD::FDIV, VT, Custom); 1473 setOperationAction(ISD::FFLOOR, VT, Custom); 1474 setOperationAction(ISD::FMA, VT, Custom); 1475 setOperationAction(ISD::FMAXIMUM, VT, Custom); 1476 setOperationAction(ISD::FMAXNUM, VT, Custom); 1477 setOperationAction(ISD::FMINIMUM, VT, Custom); 1478 setOperationAction(ISD::FMINNUM, VT, Custom); 1479 setOperationAction(ISD::FMUL, VT, Custom); 1480 setOperationAction(ISD::FNEARBYINT, VT, Custom); 1481 setOperationAction(ISD::FNEG, VT, Custom); 1482 setOperationAction(ISD::FRINT, VT, Custom); 1483 setOperationAction(ISD::FROUND, VT, Custom); 1484 setOperationAction(ISD::FROUNDEVEN, VT, Custom); 1485 setOperationAction(ISD::FSQRT, VT, Custom); 1486 setOperationAction(ISD::FSUB, VT, Custom); 1487 setOperationAction(ISD::FTRUNC, VT, Custom); 1488 setOperationAction(ISD::LOAD, VT, Custom); 1489 setOperationAction(ISD::MLOAD, VT, Custom); 1490 setOperationAction(ISD::MSTORE, VT, Custom); 1491 setOperationAction(ISD::MUL, VT, Custom); 1492 setOperationAction(ISD::MULHS, VT, Custom); 1493 setOperationAction(ISD::MULHU, VT, Custom); 1494 setOperationAction(ISD::OR, VT, Custom); 1495 setOperationAction(ISD::SDIV, VT, Custom); 1496 setOperationAction(ISD::SELECT, VT, Custom); 1497 setOperationAction(ISD::SETCC, VT, Custom); 1498 setOperationAction(ISD::SHL, VT, Custom); 1499 setOperationAction(ISD::SIGN_EXTEND, VT, Custom); 1500 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Custom); 1501 setOperationAction(ISD::SMAX, VT, Custom); 1502 setOperationAction(ISD::SMIN, VT, Custom); 1503 setOperationAction(ISD::SPLAT_VECTOR, VT, Custom); 1504 setOperationAction(ISD::SRA, VT, Custom); 1505 setOperationAction(ISD::SRL, VT, Custom); 1506 setOperationAction(ISD::STORE, VT, Custom); 1507 setOperationAction(ISD::SUB, VT, Custom); 1508 setOperationAction(ISD::TRUNCATE, VT, Custom); 1509 setOperationAction(ISD::UDIV, VT, Custom); 1510 setOperationAction(ISD::UMAX, VT, Custom); 1511 setOperationAction(ISD::UMIN, VT, Custom); 1512 setOperationAction(ISD::VECREDUCE_ADD, VT, Custom); 1513 setOperationAction(ISD::VECREDUCE_AND, VT, Custom); 1514 setOperationAction(ISD::VECREDUCE_FADD, VT, Custom); 1515 setOperationAction(ISD::VECREDUCE_SEQ_FADD, VT, Custom); 1516 setOperationAction(ISD::VECREDUCE_FMAX, VT, Custom); 1517 setOperationAction(ISD::VECREDUCE_FMIN, VT, Custom); 1518 setOperationAction(ISD::VECREDUCE_OR, VT, Custom); 1519 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom); 1520 setOperationAction(ISD::VECREDUCE_SMAX, VT, Custom); 1521 setOperationAction(ISD::VECREDUCE_SMIN, VT, Custom); 1522 setOperationAction(ISD::VECREDUCE_UMAX, VT, Custom); 1523 setOperationAction(ISD::VECREDUCE_UMIN, VT, Custom); 1524 setOperationAction(ISD::VECREDUCE_XOR, VT, Custom); 1525 setOperationAction(ISD::VSELECT, VT, Custom); 1526 setOperationAction(ISD::XOR, VT, Custom); 1527 setOperationAction(ISD::ZERO_EXTEND, VT, Custom); 1528 } 1529 1530 void AArch64TargetLowering::addDRTypeForNEON(MVT VT) { 1531 addRegisterClass(VT, &AArch64::FPR64RegClass); 1532 addTypeForNEON(VT, MVT::v2i32); 1533 } 1534 1535 void AArch64TargetLowering::addQRTypeForNEON(MVT VT) { 1536 addRegisterClass(VT, &AArch64::FPR128RegClass); 1537 addTypeForNEON(VT, MVT::v4i32); 1538 } 1539 1540 EVT AArch64TargetLowering::getSetCCResultType(const DataLayout &, 1541 LLVMContext &C, EVT VT) const { 1542 if (!VT.isVector()) 1543 return MVT::i32; 1544 if (VT.isScalableVector()) 1545 return EVT::getVectorVT(C, MVT::i1, VT.getVectorElementCount()); 1546 return VT.changeVectorElementTypeToInteger(); 1547 } 1548 1549 static bool optimizeLogicalImm(SDValue Op, unsigned Size, uint64_t Imm, 1550 const APInt &Demanded, 1551 TargetLowering::TargetLoweringOpt &TLO, 1552 unsigned NewOpc) { 1553 uint64_t OldImm = Imm, NewImm, Enc; 1554 uint64_t Mask = ((uint64_t)(-1LL) >> (64 - Size)), OrigMask = Mask; 1555 1556 // Return if the immediate is already all zeros, all ones, a bimm32 or a 1557 // bimm64. 1558 if (Imm == 0 || Imm == Mask || 1559 AArch64_AM::isLogicalImmediate(Imm & Mask, Size)) 1560 return false; 1561 1562 unsigned EltSize = Size; 1563 uint64_t DemandedBits = Demanded.getZExtValue(); 1564 1565 // Clear bits that are not demanded. 1566 Imm &= DemandedBits; 1567 1568 while (true) { 1569 // The goal here is to set the non-demanded bits in a way that minimizes 1570 // the number of switching between 0 and 1. In order to achieve this goal, 1571 // we set the non-demanded bits to the value of the preceding demanded bits. 1572 // For example, if we have an immediate 0bx10xx0x1 ('x' indicates a 1573 // non-demanded bit), we copy bit0 (1) to the least significant 'x', 1574 // bit2 (0) to 'xx', and bit6 (1) to the most significant 'x'. 1575 // The final result is 0b11000011. 1576 uint64_t NonDemandedBits = ~DemandedBits; 1577 uint64_t InvertedImm = ~Imm & DemandedBits; 1578 uint64_t RotatedImm = 1579 ((InvertedImm << 1) | (InvertedImm >> (EltSize - 1) & 1)) & 1580 NonDemandedBits; 1581 uint64_t Sum = RotatedImm + NonDemandedBits; 1582 bool Carry = NonDemandedBits & ~Sum & (1ULL << (EltSize - 1)); 1583 uint64_t Ones = (Sum + Carry) & NonDemandedBits; 1584 NewImm = (Imm | Ones) & Mask; 1585 1586 // If NewImm or its bitwise NOT is a shifted mask, it is a bitmask immediate 1587 // or all-ones or all-zeros, in which case we can stop searching. Otherwise, 1588 // we halve the element size and continue the search. 1589 if (isShiftedMask_64(NewImm) || isShiftedMask_64(~(NewImm | ~Mask))) 1590 break; 1591 1592 // We cannot shrink the element size any further if it is 2-bits. 1593 if (EltSize == 2) 1594 return false; 1595 1596 EltSize /= 2; 1597 Mask >>= EltSize; 1598 uint64_t Hi = Imm >> EltSize, DemandedBitsHi = DemandedBits >> EltSize; 1599 1600 // Return if there is mismatch in any of the demanded bits of Imm and Hi. 1601 if (((Imm ^ Hi) & (DemandedBits & DemandedBitsHi) & Mask) != 0) 1602 return false; 1603 1604 // Merge the upper and lower halves of Imm and DemandedBits. 1605 Imm |= Hi; 1606 DemandedBits |= DemandedBitsHi; 1607 } 1608 1609 ++NumOptimizedImms; 1610 1611 // Replicate the element across the register width. 1612 while (EltSize < Size) { 1613 NewImm |= NewImm << EltSize; 1614 EltSize *= 2; 1615 } 1616 1617 (void)OldImm; 1618 assert(((OldImm ^ NewImm) & Demanded.getZExtValue()) == 0 && 1619 "demanded bits should never be altered"); 1620 assert(OldImm != NewImm && "the new imm shouldn't be equal to the old imm"); 1621 1622 // Create the new constant immediate node. 1623 EVT VT = Op.getValueType(); 1624 SDLoc DL(Op); 1625 SDValue New; 1626 1627 // If the new constant immediate is all-zeros or all-ones, let the target 1628 // independent DAG combine optimize this node. 1629 if (NewImm == 0 || NewImm == OrigMask) { 1630 New = TLO.DAG.getNode(Op.getOpcode(), DL, VT, Op.getOperand(0), 1631 TLO.DAG.getConstant(NewImm, DL, VT)); 1632 // Otherwise, create a machine node so that target independent DAG combine 1633 // doesn't undo this optimization. 1634 } else { 1635 Enc = AArch64_AM::encodeLogicalImmediate(NewImm, Size); 1636 SDValue EncConst = TLO.DAG.getTargetConstant(Enc, DL, VT); 1637 New = SDValue( 1638 TLO.DAG.getMachineNode(NewOpc, DL, VT, Op.getOperand(0), EncConst), 0); 1639 } 1640 1641 return TLO.CombineTo(Op, New); 1642 } 1643 1644 bool AArch64TargetLowering::targetShrinkDemandedConstant( 1645 SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts, 1646 TargetLoweringOpt &TLO) const { 1647 // Delay this optimization to as late as possible. 1648 if (!TLO.LegalOps) 1649 return false; 1650 1651 if (!EnableOptimizeLogicalImm) 1652 return false; 1653 1654 EVT VT = Op.getValueType(); 1655 if (VT.isVector()) 1656 return false; 1657 1658 unsigned Size = VT.getSizeInBits(); 1659 assert((Size == 32 || Size == 64) && 1660 "i32 or i64 is expected after legalization."); 1661 1662 // Exit early if we demand all bits. 1663 if (DemandedBits.countPopulation() == Size) 1664 return false; 1665 1666 unsigned NewOpc; 1667 switch (Op.getOpcode()) { 1668 default: 1669 return false; 1670 case ISD::AND: 1671 NewOpc = Size == 32 ? AArch64::ANDWri : AArch64::ANDXri; 1672 break; 1673 case ISD::OR: 1674 NewOpc = Size == 32 ? AArch64::ORRWri : AArch64::ORRXri; 1675 break; 1676 case ISD::XOR: 1677 NewOpc = Size == 32 ? AArch64::EORWri : AArch64::EORXri; 1678 break; 1679 } 1680 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)); 1681 if (!C) 1682 return false; 1683 uint64_t Imm = C->getZExtValue(); 1684 return optimizeLogicalImm(Op, Size, Imm, DemandedBits, TLO, NewOpc); 1685 } 1686 1687 /// computeKnownBitsForTargetNode - Determine which of the bits specified in 1688 /// Mask are known to be either zero or one and return them Known. 1689 void AArch64TargetLowering::computeKnownBitsForTargetNode( 1690 const SDValue Op, KnownBits &Known, 1691 const APInt &DemandedElts, const SelectionDAG &DAG, unsigned Depth) const { 1692 switch (Op.getOpcode()) { 1693 default: 1694 break; 1695 case AArch64ISD::CSEL: { 1696 KnownBits Known2; 1697 Known = DAG.computeKnownBits(Op->getOperand(0), Depth + 1); 1698 Known2 = DAG.computeKnownBits(Op->getOperand(1), Depth + 1); 1699 Known = KnownBits::commonBits(Known, Known2); 1700 break; 1701 } 1702 case AArch64ISD::LOADgot: 1703 case AArch64ISD::ADDlow: { 1704 if (!Subtarget->isTargetILP32()) 1705 break; 1706 // In ILP32 mode all valid pointers are in the low 4GB of the address-space. 1707 Known.Zero = APInt::getHighBitsSet(64, 32); 1708 break; 1709 } 1710 case ISD::INTRINSIC_W_CHAIN: { 1711 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1)); 1712 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue()); 1713 switch (IntID) { 1714 default: return; 1715 case Intrinsic::aarch64_ldaxr: 1716 case Intrinsic::aarch64_ldxr: { 1717 unsigned BitWidth = Known.getBitWidth(); 1718 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT(); 1719 unsigned MemBits = VT.getScalarSizeInBits(); 1720 Known.Zero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits); 1721 return; 1722 } 1723 } 1724 break; 1725 } 1726 case ISD::INTRINSIC_WO_CHAIN: 1727 case ISD::INTRINSIC_VOID: { 1728 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 1729 switch (IntNo) { 1730 default: 1731 break; 1732 case Intrinsic::aarch64_neon_umaxv: 1733 case Intrinsic::aarch64_neon_uminv: { 1734 // Figure out the datatype of the vector operand. The UMINV instruction 1735 // will zero extend the result, so we can mark as known zero all the 1736 // bits larger than the element datatype. 32-bit or larget doesn't need 1737 // this as those are legal types and will be handled by isel directly. 1738 MVT VT = Op.getOperand(1).getValueType().getSimpleVT(); 1739 unsigned BitWidth = Known.getBitWidth(); 1740 if (VT == MVT::v8i8 || VT == MVT::v16i8) { 1741 assert(BitWidth >= 8 && "Unexpected width!"); 1742 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 8); 1743 Known.Zero |= Mask; 1744 } else if (VT == MVT::v4i16 || VT == MVT::v8i16) { 1745 assert(BitWidth >= 16 && "Unexpected width!"); 1746 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 16); 1747 Known.Zero |= Mask; 1748 } 1749 break; 1750 } break; 1751 } 1752 } 1753 } 1754 } 1755 1756 MVT AArch64TargetLowering::getScalarShiftAmountTy(const DataLayout &DL, 1757 EVT) const { 1758 return MVT::i64; 1759 } 1760 1761 bool AArch64TargetLowering::allowsMisalignedMemoryAccesses( 1762 EVT VT, unsigned AddrSpace, Align Alignment, MachineMemOperand::Flags Flags, 1763 bool *Fast) const { 1764 if (Subtarget->requiresStrictAlign()) 1765 return false; 1766 1767 if (Fast) { 1768 // Some CPUs are fine with unaligned stores except for 128-bit ones. 1769 *Fast = !Subtarget->isMisaligned128StoreSlow() || VT.getStoreSize() != 16 || 1770 // See comments in performSTORECombine() for more details about 1771 // these conditions. 1772 1773 // Code that uses clang vector extensions can mark that it 1774 // wants unaligned accesses to be treated as fast by 1775 // underspecifying alignment to be 1 or 2. 1776 Alignment <= 2 || 1777 1778 // Disregard v2i64. Memcpy lowering produces those and splitting 1779 // them regresses performance on micro-benchmarks and olden/bh. 1780 VT == MVT::v2i64; 1781 } 1782 return true; 1783 } 1784 1785 // Same as above but handling LLTs instead. 1786 bool AArch64TargetLowering::allowsMisalignedMemoryAccesses( 1787 LLT Ty, unsigned AddrSpace, Align Alignment, MachineMemOperand::Flags Flags, 1788 bool *Fast) const { 1789 if (Subtarget->requiresStrictAlign()) 1790 return false; 1791 1792 if (Fast) { 1793 // Some CPUs are fine with unaligned stores except for 128-bit ones. 1794 *Fast = !Subtarget->isMisaligned128StoreSlow() || 1795 Ty.getSizeInBytes() != 16 || 1796 // See comments in performSTORECombine() for more details about 1797 // these conditions. 1798 1799 // Code that uses clang vector extensions can mark that it 1800 // wants unaligned accesses to be treated as fast by 1801 // underspecifying alignment to be 1 or 2. 1802 Alignment <= 2 || 1803 1804 // Disregard v2i64. Memcpy lowering produces those and splitting 1805 // them regresses performance on micro-benchmarks and olden/bh. 1806 Ty == LLT::vector(2, 64); 1807 } 1808 return true; 1809 } 1810 1811 FastISel * 1812 AArch64TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo, 1813 const TargetLibraryInfo *libInfo) const { 1814 return AArch64::createFastISel(funcInfo, libInfo); 1815 } 1816 1817 const char *AArch64TargetLowering::getTargetNodeName(unsigned Opcode) const { 1818 #define MAKE_CASE(V) \ 1819 case V: \ 1820 return #V; 1821 switch ((AArch64ISD::NodeType)Opcode) { 1822 case AArch64ISD::FIRST_NUMBER: 1823 break; 1824 MAKE_CASE(AArch64ISD::CALL) 1825 MAKE_CASE(AArch64ISD::ADRP) 1826 MAKE_CASE(AArch64ISD::ADR) 1827 MAKE_CASE(AArch64ISD::ADDlow) 1828 MAKE_CASE(AArch64ISD::LOADgot) 1829 MAKE_CASE(AArch64ISD::RET_FLAG) 1830 MAKE_CASE(AArch64ISD::BRCOND) 1831 MAKE_CASE(AArch64ISD::CSEL) 1832 MAKE_CASE(AArch64ISD::CSINV) 1833 MAKE_CASE(AArch64ISD::CSNEG) 1834 MAKE_CASE(AArch64ISD::CSINC) 1835 MAKE_CASE(AArch64ISD::THREAD_POINTER) 1836 MAKE_CASE(AArch64ISD::TLSDESC_CALLSEQ) 1837 MAKE_CASE(AArch64ISD::ADD_PRED) 1838 MAKE_CASE(AArch64ISD::MUL_PRED) 1839 MAKE_CASE(AArch64ISD::MULHS_PRED) 1840 MAKE_CASE(AArch64ISD::MULHU_PRED) 1841 MAKE_CASE(AArch64ISD::SDIV_PRED) 1842 MAKE_CASE(AArch64ISD::SHL_PRED) 1843 MAKE_CASE(AArch64ISD::SMAX_PRED) 1844 MAKE_CASE(AArch64ISD::SMIN_PRED) 1845 MAKE_CASE(AArch64ISD::SRA_PRED) 1846 MAKE_CASE(AArch64ISD::SRL_PRED) 1847 MAKE_CASE(AArch64ISD::SUB_PRED) 1848 MAKE_CASE(AArch64ISD::UDIV_PRED) 1849 MAKE_CASE(AArch64ISD::UMAX_PRED) 1850 MAKE_CASE(AArch64ISD::UMIN_PRED) 1851 MAKE_CASE(AArch64ISD::FNEG_MERGE_PASSTHRU) 1852 MAKE_CASE(AArch64ISD::SIGN_EXTEND_INREG_MERGE_PASSTHRU) 1853 MAKE_CASE(AArch64ISD::ZERO_EXTEND_INREG_MERGE_PASSTHRU) 1854 MAKE_CASE(AArch64ISD::FCEIL_MERGE_PASSTHRU) 1855 MAKE_CASE(AArch64ISD::FFLOOR_MERGE_PASSTHRU) 1856 MAKE_CASE(AArch64ISD::FNEARBYINT_MERGE_PASSTHRU) 1857 MAKE_CASE(AArch64ISD::FRINT_MERGE_PASSTHRU) 1858 MAKE_CASE(AArch64ISD::FROUND_MERGE_PASSTHRU) 1859 MAKE_CASE(AArch64ISD::FROUNDEVEN_MERGE_PASSTHRU) 1860 MAKE_CASE(AArch64ISD::FTRUNC_MERGE_PASSTHRU) 1861 MAKE_CASE(AArch64ISD::FP_ROUND_MERGE_PASSTHRU) 1862 MAKE_CASE(AArch64ISD::FP_EXTEND_MERGE_PASSTHRU) 1863 MAKE_CASE(AArch64ISD::SINT_TO_FP_MERGE_PASSTHRU) 1864 MAKE_CASE(AArch64ISD::UINT_TO_FP_MERGE_PASSTHRU) 1865 MAKE_CASE(AArch64ISD::FCVTZU_MERGE_PASSTHRU) 1866 MAKE_CASE(AArch64ISD::FCVTZS_MERGE_PASSTHRU) 1867 MAKE_CASE(AArch64ISD::FSQRT_MERGE_PASSTHRU) 1868 MAKE_CASE(AArch64ISD::FRECPX_MERGE_PASSTHRU) 1869 MAKE_CASE(AArch64ISD::FABS_MERGE_PASSTHRU) 1870 MAKE_CASE(AArch64ISD::ABS_MERGE_PASSTHRU) 1871 MAKE_CASE(AArch64ISD::NEG_MERGE_PASSTHRU) 1872 MAKE_CASE(AArch64ISD::SETCC_MERGE_ZERO) 1873 MAKE_CASE(AArch64ISD::ADC) 1874 MAKE_CASE(AArch64ISD::SBC) 1875 MAKE_CASE(AArch64ISD::ADDS) 1876 MAKE_CASE(AArch64ISD::SUBS) 1877 MAKE_CASE(AArch64ISD::ADCS) 1878 MAKE_CASE(AArch64ISD::SBCS) 1879 MAKE_CASE(AArch64ISD::ANDS) 1880 MAKE_CASE(AArch64ISD::CCMP) 1881 MAKE_CASE(AArch64ISD::CCMN) 1882 MAKE_CASE(AArch64ISD::FCCMP) 1883 MAKE_CASE(AArch64ISD::FCMP) 1884 MAKE_CASE(AArch64ISD::STRICT_FCMP) 1885 MAKE_CASE(AArch64ISD::STRICT_FCMPE) 1886 MAKE_CASE(AArch64ISD::DUP) 1887 MAKE_CASE(AArch64ISD::DUPLANE8) 1888 MAKE_CASE(AArch64ISD::DUPLANE16) 1889 MAKE_CASE(AArch64ISD::DUPLANE32) 1890 MAKE_CASE(AArch64ISD::DUPLANE64) 1891 MAKE_CASE(AArch64ISD::MOVI) 1892 MAKE_CASE(AArch64ISD::MOVIshift) 1893 MAKE_CASE(AArch64ISD::MOVIedit) 1894 MAKE_CASE(AArch64ISD::MOVImsl) 1895 MAKE_CASE(AArch64ISD::FMOV) 1896 MAKE_CASE(AArch64ISD::MVNIshift) 1897 MAKE_CASE(AArch64ISD::MVNImsl) 1898 MAKE_CASE(AArch64ISD::BICi) 1899 MAKE_CASE(AArch64ISD::ORRi) 1900 MAKE_CASE(AArch64ISD::BSP) 1901 MAKE_CASE(AArch64ISD::NEG) 1902 MAKE_CASE(AArch64ISD::EXTR) 1903 MAKE_CASE(AArch64ISD::ZIP1) 1904 MAKE_CASE(AArch64ISD::ZIP2) 1905 MAKE_CASE(AArch64ISD::UZP1) 1906 MAKE_CASE(AArch64ISD::UZP2) 1907 MAKE_CASE(AArch64ISD::TRN1) 1908 MAKE_CASE(AArch64ISD::TRN2) 1909 MAKE_CASE(AArch64ISD::REV16) 1910 MAKE_CASE(AArch64ISD::REV32) 1911 MAKE_CASE(AArch64ISD::REV64) 1912 MAKE_CASE(AArch64ISD::EXT) 1913 MAKE_CASE(AArch64ISD::VSHL) 1914 MAKE_CASE(AArch64ISD::VLSHR) 1915 MAKE_CASE(AArch64ISD::VASHR) 1916 MAKE_CASE(AArch64ISD::VSLI) 1917 MAKE_CASE(AArch64ISD::VSRI) 1918 MAKE_CASE(AArch64ISD::CMEQ) 1919 MAKE_CASE(AArch64ISD::CMGE) 1920 MAKE_CASE(AArch64ISD::CMGT) 1921 MAKE_CASE(AArch64ISD::CMHI) 1922 MAKE_CASE(AArch64ISD::CMHS) 1923 MAKE_CASE(AArch64ISD::FCMEQ) 1924 MAKE_CASE(AArch64ISD::FCMGE) 1925 MAKE_CASE(AArch64ISD::FCMGT) 1926 MAKE_CASE(AArch64ISD::CMEQz) 1927 MAKE_CASE(AArch64ISD::CMGEz) 1928 MAKE_CASE(AArch64ISD::CMGTz) 1929 MAKE_CASE(AArch64ISD::CMLEz) 1930 MAKE_CASE(AArch64ISD::CMLTz) 1931 MAKE_CASE(AArch64ISD::FCMEQz) 1932 MAKE_CASE(AArch64ISD::FCMGEz) 1933 MAKE_CASE(AArch64ISD::FCMGTz) 1934 MAKE_CASE(AArch64ISD::FCMLEz) 1935 MAKE_CASE(AArch64ISD::FCMLTz) 1936 MAKE_CASE(AArch64ISD::SADDV) 1937 MAKE_CASE(AArch64ISD::UADDV) 1938 MAKE_CASE(AArch64ISD::SRHADD) 1939 MAKE_CASE(AArch64ISD::URHADD) 1940 MAKE_CASE(AArch64ISD::SHADD) 1941 MAKE_CASE(AArch64ISD::UHADD) 1942 MAKE_CASE(AArch64ISD::SDOT) 1943 MAKE_CASE(AArch64ISD::UDOT) 1944 MAKE_CASE(AArch64ISD::SMINV) 1945 MAKE_CASE(AArch64ISD::UMINV) 1946 MAKE_CASE(AArch64ISD::SMAXV) 1947 MAKE_CASE(AArch64ISD::UMAXV) 1948 MAKE_CASE(AArch64ISD::SADDV_PRED) 1949 MAKE_CASE(AArch64ISD::UADDV_PRED) 1950 MAKE_CASE(AArch64ISD::SMAXV_PRED) 1951 MAKE_CASE(AArch64ISD::UMAXV_PRED) 1952 MAKE_CASE(AArch64ISD::SMINV_PRED) 1953 MAKE_CASE(AArch64ISD::UMINV_PRED) 1954 MAKE_CASE(AArch64ISD::ORV_PRED) 1955 MAKE_CASE(AArch64ISD::EORV_PRED) 1956 MAKE_CASE(AArch64ISD::ANDV_PRED) 1957 MAKE_CASE(AArch64ISD::CLASTA_N) 1958 MAKE_CASE(AArch64ISD::CLASTB_N) 1959 MAKE_CASE(AArch64ISD::LASTA) 1960 MAKE_CASE(AArch64ISD::LASTB) 1961 MAKE_CASE(AArch64ISD::REINTERPRET_CAST) 1962 MAKE_CASE(AArch64ISD::TBL) 1963 MAKE_CASE(AArch64ISD::FADD_PRED) 1964 MAKE_CASE(AArch64ISD::FADDA_PRED) 1965 MAKE_CASE(AArch64ISD::FADDV_PRED) 1966 MAKE_CASE(AArch64ISD::FDIV_PRED) 1967 MAKE_CASE(AArch64ISD::FMA_PRED) 1968 MAKE_CASE(AArch64ISD::FMAX_PRED) 1969 MAKE_CASE(AArch64ISD::FMAXV_PRED) 1970 MAKE_CASE(AArch64ISD::FMAXNM_PRED) 1971 MAKE_CASE(AArch64ISD::FMAXNMV_PRED) 1972 MAKE_CASE(AArch64ISD::FMIN_PRED) 1973 MAKE_CASE(AArch64ISD::FMINV_PRED) 1974 MAKE_CASE(AArch64ISD::FMINNM_PRED) 1975 MAKE_CASE(AArch64ISD::FMINNMV_PRED) 1976 MAKE_CASE(AArch64ISD::FMUL_PRED) 1977 MAKE_CASE(AArch64ISD::FSUB_PRED) 1978 MAKE_CASE(AArch64ISD::BIC) 1979 MAKE_CASE(AArch64ISD::BIT) 1980 MAKE_CASE(AArch64ISD::CBZ) 1981 MAKE_CASE(AArch64ISD::CBNZ) 1982 MAKE_CASE(AArch64ISD::TBZ) 1983 MAKE_CASE(AArch64ISD::TBNZ) 1984 MAKE_CASE(AArch64ISD::TC_RETURN) 1985 MAKE_CASE(AArch64ISD::PREFETCH) 1986 MAKE_CASE(AArch64ISD::SITOF) 1987 MAKE_CASE(AArch64ISD::UITOF) 1988 MAKE_CASE(AArch64ISD::NVCAST) 1989 MAKE_CASE(AArch64ISD::MRS) 1990 MAKE_CASE(AArch64ISD::SQSHL_I) 1991 MAKE_CASE(AArch64ISD::UQSHL_I) 1992 MAKE_CASE(AArch64ISD::SRSHR_I) 1993 MAKE_CASE(AArch64ISD::URSHR_I) 1994 MAKE_CASE(AArch64ISD::SQSHLU_I) 1995 MAKE_CASE(AArch64ISD::WrapperLarge) 1996 MAKE_CASE(AArch64ISD::LD2post) 1997 MAKE_CASE(AArch64ISD::LD3post) 1998 MAKE_CASE(AArch64ISD::LD4post) 1999 MAKE_CASE(AArch64ISD::ST2post) 2000 MAKE_CASE(AArch64ISD::ST3post) 2001 MAKE_CASE(AArch64ISD::ST4post) 2002 MAKE_CASE(AArch64ISD::LD1x2post) 2003 MAKE_CASE(AArch64ISD::LD1x3post) 2004 MAKE_CASE(AArch64ISD::LD1x4post) 2005 MAKE_CASE(AArch64ISD::ST1x2post) 2006 MAKE_CASE(AArch64ISD::ST1x3post) 2007 MAKE_CASE(AArch64ISD::ST1x4post) 2008 MAKE_CASE(AArch64ISD::LD1DUPpost) 2009 MAKE_CASE(AArch64ISD::LD2DUPpost) 2010 MAKE_CASE(AArch64ISD::LD3DUPpost) 2011 MAKE_CASE(AArch64ISD::LD4DUPpost) 2012 MAKE_CASE(AArch64ISD::LD1LANEpost) 2013 MAKE_CASE(AArch64ISD::LD2LANEpost) 2014 MAKE_CASE(AArch64ISD::LD3LANEpost) 2015 MAKE_CASE(AArch64ISD::LD4LANEpost) 2016 MAKE_CASE(AArch64ISD::ST2LANEpost) 2017 MAKE_CASE(AArch64ISD::ST3LANEpost) 2018 MAKE_CASE(AArch64ISD::ST4LANEpost) 2019 MAKE_CASE(AArch64ISD::SMULL) 2020 MAKE_CASE(AArch64ISD::UMULL) 2021 MAKE_CASE(AArch64ISD::FRECPE) 2022 MAKE_CASE(AArch64ISD::FRECPS) 2023 MAKE_CASE(AArch64ISD::FRSQRTE) 2024 MAKE_CASE(AArch64ISD::FRSQRTS) 2025 MAKE_CASE(AArch64ISD::STG) 2026 MAKE_CASE(AArch64ISD::STZG) 2027 MAKE_CASE(AArch64ISD::ST2G) 2028 MAKE_CASE(AArch64ISD::STZ2G) 2029 MAKE_CASE(AArch64ISD::SUNPKHI) 2030 MAKE_CASE(AArch64ISD::SUNPKLO) 2031 MAKE_CASE(AArch64ISD::UUNPKHI) 2032 MAKE_CASE(AArch64ISD::UUNPKLO) 2033 MAKE_CASE(AArch64ISD::INSR) 2034 MAKE_CASE(AArch64ISD::PTEST) 2035 MAKE_CASE(AArch64ISD::PTRUE) 2036 MAKE_CASE(AArch64ISD::LD1_MERGE_ZERO) 2037 MAKE_CASE(AArch64ISD::LD1S_MERGE_ZERO) 2038 MAKE_CASE(AArch64ISD::LDNF1_MERGE_ZERO) 2039 MAKE_CASE(AArch64ISD::LDNF1S_MERGE_ZERO) 2040 MAKE_CASE(AArch64ISD::LDFF1_MERGE_ZERO) 2041 MAKE_CASE(AArch64ISD::LDFF1S_MERGE_ZERO) 2042 MAKE_CASE(AArch64ISD::LD1RQ_MERGE_ZERO) 2043 MAKE_CASE(AArch64ISD::LD1RO_MERGE_ZERO) 2044 MAKE_CASE(AArch64ISD::SVE_LD2_MERGE_ZERO) 2045 MAKE_CASE(AArch64ISD::SVE_LD3_MERGE_ZERO) 2046 MAKE_CASE(AArch64ISD::SVE_LD4_MERGE_ZERO) 2047 MAKE_CASE(AArch64ISD::GLD1_MERGE_ZERO) 2048 MAKE_CASE(AArch64ISD::GLD1_SCALED_MERGE_ZERO) 2049 MAKE_CASE(AArch64ISD::GLD1_SXTW_MERGE_ZERO) 2050 MAKE_CASE(AArch64ISD::GLD1_UXTW_MERGE_ZERO) 2051 MAKE_CASE(AArch64ISD::GLD1_SXTW_SCALED_MERGE_ZERO) 2052 MAKE_CASE(AArch64ISD::GLD1_UXTW_SCALED_MERGE_ZERO) 2053 MAKE_CASE(AArch64ISD::GLD1_IMM_MERGE_ZERO) 2054 MAKE_CASE(AArch64ISD::GLD1S_MERGE_ZERO) 2055 MAKE_CASE(AArch64ISD::GLD1S_SCALED_MERGE_ZERO) 2056 MAKE_CASE(AArch64ISD::GLD1S_SXTW_MERGE_ZERO) 2057 MAKE_CASE(AArch64ISD::GLD1S_UXTW_MERGE_ZERO) 2058 MAKE_CASE(AArch64ISD::GLD1S_SXTW_SCALED_MERGE_ZERO) 2059 MAKE_CASE(AArch64ISD::GLD1S_UXTW_SCALED_MERGE_ZERO) 2060 MAKE_CASE(AArch64ISD::GLD1S_IMM_MERGE_ZERO) 2061 MAKE_CASE(AArch64ISD::GLDFF1_MERGE_ZERO) 2062 MAKE_CASE(AArch64ISD::GLDFF1_SCALED_MERGE_ZERO) 2063 MAKE_CASE(AArch64ISD::GLDFF1_SXTW_MERGE_ZERO) 2064 MAKE_CASE(AArch64ISD::GLDFF1_UXTW_MERGE_ZERO) 2065 MAKE_CASE(AArch64ISD::GLDFF1_SXTW_SCALED_MERGE_ZERO) 2066 MAKE_CASE(AArch64ISD::GLDFF1_UXTW_SCALED_MERGE_ZERO) 2067 MAKE_CASE(AArch64ISD::GLDFF1_IMM_MERGE_ZERO) 2068 MAKE_CASE(AArch64ISD::GLDFF1S_MERGE_ZERO) 2069 MAKE_CASE(AArch64ISD::GLDFF1S_SCALED_MERGE_ZERO) 2070 MAKE_CASE(AArch64ISD::GLDFF1S_SXTW_MERGE_ZERO) 2071 MAKE_CASE(AArch64ISD::GLDFF1S_UXTW_MERGE_ZERO) 2072 MAKE_CASE(AArch64ISD::GLDFF1S_SXTW_SCALED_MERGE_ZERO) 2073 MAKE_CASE(AArch64ISD::GLDFF1S_UXTW_SCALED_MERGE_ZERO) 2074 MAKE_CASE(AArch64ISD::GLDFF1S_IMM_MERGE_ZERO) 2075 MAKE_CASE(AArch64ISD::GLDNT1_MERGE_ZERO) 2076 MAKE_CASE(AArch64ISD::GLDNT1_INDEX_MERGE_ZERO) 2077 MAKE_CASE(AArch64ISD::GLDNT1S_MERGE_ZERO) 2078 MAKE_CASE(AArch64ISD::ST1_PRED) 2079 MAKE_CASE(AArch64ISD::SST1_PRED) 2080 MAKE_CASE(AArch64ISD::SST1_SCALED_PRED) 2081 MAKE_CASE(AArch64ISD::SST1_SXTW_PRED) 2082 MAKE_CASE(AArch64ISD::SST1_UXTW_PRED) 2083 MAKE_CASE(AArch64ISD::SST1_SXTW_SCALED_PRED) 2084 MAKE_CASE(AArch64ISD::SST1_UXTW_SCALED_PRED) 2085 MAKE_CASE(AArch64ISD::SST1_IMM_PRED) 2086 MAKE_CASE(AArch64ISD::SSTNT1_PRED) 2087 MAKE_CASE(AArch64ISD::SSTNT1_INDEX_PRED) 2088 MAKE_CASE(AArch64ISD::LDP) 2089 MAKE_CASE(AArch64ISD::STP) 2090 MAKE_CASE(AArch64ISD::STNP) 2091 MAKE_CASE(AArch64ISD::BITREVERSE_MERGE_PASSTHRU) 2092 MAKE_CASE(AArch64ISD::BSWAP_MERGE_PASSTHRU) 2093 MAKE_CASE(AArch64ISD::CTLZ_MERGE_PASSTHRU) 2094 MAKE_CASE(AArch64ISD::CTPOP_MERGE_PASSTHRU) 2095 MAKE_CASE(AArch64ISD::DUP_MERGE_PASSTHRU) 2096 MAKE_CASE(AArch64ISD::INDEX_VECTOR) 2097 MAKE_CASE(AArch64ISD::UABD) 2098 MAKE_CASE(AArch64ISD::SABD) 2099 MAKE_CASE(AArch64ISD::CALL_RVMARKER) 2100 } 2101 #undef MAKE_CASE 2102 return nullptr; 2103 } 2104 2105 MachineBasicBlock * 2106 AArch64TargetLowering::EmitF128CSEL(MachineInstr &MI, 2107 MachineBasicBlock *MBB) const { 2108 // We materialise the F128CSEL pseudo-instruction as some control flow and a 2109 // phi node: 2110 2111 // OrigBB: 2112 // [... previous instrs leading to comparison ...] 2113 // b.ne TrueBB 2114 // b EndBB 2115 // TrueBB: 2116 // ; Fallthrough 2117 // EndBB: 2118 // Dest = PHI [IfTrue, TrueBB], [IfFalse, OrigBB] 2119 2120 MachineFunction *MF = MBB->getParent(); 2121 const TargetInstrInfo *TII = Subtarget->getInstrInfo(); 2122 const BasicBlock *LLVM_BB = MBB->getBasicBlock(); 2123 DebugLoc DL = MI.getDebugLoc(); 2124 MachineFunction::iterator It = ++MBB->getIterator(); 2125 2126 Register DestReg = MI.getOperand(0).getReg(); 2127 Register IfTrueReg = MI.getOperand(1).getReg(); 2128 Register IfFalseReg = MI.getOperand(2).getReg(); 2129 unsigned CondCode = MI.getOperand(3).getImm(); 2130 bool NZCVKilled = MI.getOperand(4).isKill(); 2131 2132 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB); 2133 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB); 2134 MF->insert(It, TrueBB); 2135 MF->insert(It, EndBB); 2136 2137 // Transfer rest of current basic-block to EndBB 2138 EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)), 2139 MBB->end()); 2140 EndBB->transferSuccessorsAndUpdatePHIs(MBB); 2141 2142 BuildMI(MBB, DL, TII->get(AArch64::Bcc)).addImm(CondCode).addMBB(TrueBB); 2143 BuildMI(MBB, DL, TII->get(AArch64::B)).addMBB(EndBB); 2144 MBB->addSuccessor(TrueBB); 2145 MBB->addSuccessor(EndBB); 2146 2147 // TrueBB falls through to the end. 2148 TrueBB->addSuccessor(EndBB); 2149 2150 if (!NZCVKilled) { 2151 TrueBB->addLiveIn(AArch64::NZCV); 2152 EndBB->addLiveIn(AArch64::NZCV); 2153 } 2154 2155 BuildMI(*EndBB, EndBB->begin(), DL, TII->get(AArch64::PHI), DestReg) 2156 .addReg(IfTrueReg) 2157 .addMBB(TrueBB) 2158 .addReg(IfFalseReg) 2159 .addMBB(MBB); 2160 2161 MI.eraseFromParent(); 2162 return EndBB; 2163 } 2164 2165 MachineBasicBlock *AArch64TargetLowering::EmitLoweredCatchRet( 2166 MachineInstr &MI, MachineBasicBlock *BB) const { 2167 assert(!isAsynchronousEHPersonality(classifyEHPersonality( 2168 BB->getParent()->getFunction().getPersonalityFn())) && 2169 "SEH does not use catchret!"); 2170 return BB; 2171 } 2172 2173 MachineBasicBlock *AArch64TargetLowering::EmitInstrWithCustomInserter( 2174 MachineInstr &MI, MachineBasicBlock *BB) const { 2175 switch (MI.getOpcode()) { 2176 default: 2177 #ifndef NDEBUG 2178 MI.dump(); 2179 #endif 2180 llvm_unreachable("Unexpected instruction for custom inserter!"); 2181 2182 case AArch64::F128CSEL: 2183 return EmitF128CSEL(MI, BB); 2184 2185 case TargetOpcode::STACKMAP: 2186 case TargetOpcode::PATCHPOINT: 2187 case TargetOpcode::STATEPOINT: 2188 return emitPatchPoint(MI, BB); 2189 2190 case AArch64::CATCHRET: 2191 return EmitLoweredCatchRet(MI, BB); 2192 } 2193 } 2194 2195 //===----------------------------------------------------------------------===// 2196 // AArch64 Lowering private implementation. 2197 //===----------------------------------------------------------------------===// 2198 2199 //===----------------------------------------------------------------------===// 2200 // Lowering Code 2201 //===----------------------------------------------------------------------===// 2202 2203 /// isZerosVector - Check whether SDNode N is a zero-filled vector. 2204 static bool isZerosVector(const SDNode *N) { 2205 // Look through a bit convert. 2206 while (N->getOpcode() == ISD::BITCAST) 2207 N = N->getOperand(0).getNode(); 2208 2209 if (ISD::isConstantSplatVectorAllZeros(N)) 2210 return true; 2211 2212 if (N->getOpcode() != AArch64ISD::DUP) 2213 return false; 2214 2215 auto Opnd0 = N->getOperand(0); 2216 auto *CINT = dyn_cast<ConstantSDNode>(Opnd0); 2217 auto *CFP = dyn_cast<ConstantFPSDNode>(Opnd0); 2218 return (CINT && CINT->isNullValue()) || (CFP && CFP->isZero()); 2219 } 2220 2221 /// changeIntCCToAArch64CC - Convert a DAG integer condition code to an AArch64 2222 /// CC 2223 static AArch64CC::CondCode changeIntCCToAArch64CC(ISD::CondCode CC) { 2224 switch (CC) { 2225 default: 2226 llvm_unreachable("Unknown condition code!"); 2227 case ISD::SETNE: 2228 return AArch64CC::NE; 2229 case ISD::SETEQ: 2230 return AArch64CC::EQ; 2231 case ISD::SETGT: 2232 return AArch64CC::GT; 2233 case ISD::SETGE: 2234 return AArch64CC::GE; 2235 case ISD::SETLT: 2236 return AArch64CC::LT; 2237 case ISD::SETLE: 2238 return AArch64CC::LE; 2239 case ISD::SETUGT: 2240 return AArch64CC::HI; 2241 case ISD::SETUGE: 2242 return AArch64CC::HS; 2243 case ISD::SETULT: 2244 return AArch64CC::LO; 2245 case ISD::SETULE: 2246 return AArch64CC::LS; 2247 } 2248 } 2249 2250 /// changeFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64 CC. 2251 static void changeFPCCToAArch64CC(ISD::CondCode CC, 2252 AArch64CC::CondCode &CondCode, 2253 AArch64CC::CondCode &CondCode2) { 2254 CondCode2 = AArch64CC::AL; 2255 switch (CC) { 2256 default: 2257 llvm_unreachable("Unknown FP condition!"); 2258 case ISD::SETEQ: 2259 case ISD::SETOEQ: 2260 CondCode = AArch64CC::EQ; 2261 break; 2262 case ISD::SETGT: 2263 case ISD::SETOGT: 2264 CondCode = AArch64CC::GT; 2265 break; 2266 case ISD::SETGE: 2267 case ISD::SETOGE: 2268 CondCode = AArch64CC::GE; 2269 break; 2270 case ISD::SETOLT: 2271 CondCode = AArch64CC::MI; 2272 break; 2273 case ISD::SETOLE: 2274 CondCode = AArch64CC::LS; 2275 break; 2276 case ISD::SETONE: 2277 CondCode = AArch64CC::MI; 2278 CondCode2 = AArch64CC::GT; 2279 break; 2280 case ISD::SETO: 2281 CondCode = AArch64CC::VC; 2282 break; 2283 case ISD::SETUO: 2284 CondCode = AArch64CC::VS; 2285 break; 2286 case ISD::SETUEQ: 2287 CondCode = AArch64CC::EQ; 2288 CondCode2 = AArch64CC::VS; 2289 break; 2290 case ISD::SETUGT: 2291 CondCode = AArch64CC::HI; 2292 break; 2293 case ISD::SETUGE: 2294 CondCode = AArch64CC::PL; 2295 break; 2296 case ISD::SETLT: 2297 case ISD::SETULT: 2298 CondCode = AArch64CC::LT; 2299 break; 2300 case ISD::SETLE: 2301 case ISD::SETULE: 2302 CondCode = AArch64CC::LE; 2303 break; 2304 case ISD::SETNE: 2305 case ISD::SETUNE: 2306 CondCode = AArch64CC::NE; 2307 break; 2308 } 2309 } 2310 2311 /// Convert a DAG fp condition code to an AArch64 CC. 2312 /// This differs from changeFPCCToAArch64CC in that it returns cond codes that 2313 /// should be AND'ed instead of OR'ed. 2314 static void changeFPCCToANDAArch64CC(ISD::CondCode CC, 2315 AArch64CC::CondCode &CondCode, 2316 AArch64CC::CondCode &CondCode2) { 2317 CondCode2 = AArch64CC::AL; 2318 switch (CC) { 2319 default: 2320 changeFPCCToAArch64CC(CC, CondCode, CondCode2); 2321 assert(CondCode2 == AArch64CC::AL); 2322 break; 2323 case ISD::SETONE: 2324 // (a one b) 2325 // == ((a olt b) || (a ogt b)) 2326 // == ((a ord b) && (a une b)) 2327 CondCode = AArch64CC::VC; 2328 CondCode2 = AArch64CC::NE; 2329 break; 2330 case ISD::SETUEQ: 2331 // (a ueq b) 2332 // == ((a uno b) || (a oeq b)) 2333 // == ((a ule b) && (a uge b)) 2334 CondCode = AArch64CC::PL; 2335 CondCode2 = AArch64CC::LE; 2336 break; 2337 } 2338 } 2339 2340 /// changeVectorFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64 2341 /// CC usable with the vector instructions. Fewer operations are available 2342 /// without a real NZCV register, so we have to use less efficient combinations 2343 /// to get the same effect. 2344 static void changeVectorFPCCToAArch64CC(ISD::CondCode CC, 2345 AArch64CC::CondCode &CondCode, 2346 AArch64CC::CondCode &CondCode2, 2347 bool &Invert) { 2348 Invert = false; 2349 switch (CC) { 2350 default: 2351 // Mostly the scalar mappings work fine. 2352 changeFPCCToAArch64CC(CC, CondCode, CondCode2); 2353 break; 2354 case ISD::SETUO: 2355 Invert = true; 2356 LLVM_FALLTHROUGH; 2357 case ISD::SETO: 2358 CondCode = AArch64CC::MI; 2359 CondCode2 = AArch64CC::GE; 2360 break; 2361 case ISD::SETUEQ: 2362 case ISD::SETULT: 2363 case ISD::SETULE: 2364 case ISD::SETUGT: 2365 case ISD::SETUGE: 2366 // All of the compare-mask comparisons are ordered, but we can switch 2367 // between the two by a double inversion. E.g. ULE == !OGT. 2368 Invert = true; 2369 changeFPCCToAArch64CC(getSetCCInverse(CC, /* FP inverse */ MVT::f32), 2370 CondCode, CondCode2); 2371 break; 2372 } 2373 } 2374 2375 static bool isLegalArithImmed(uint64_t C) { 2376 // Matches AArch64DAGToDAGISel::SelectArithImmed(). 2377 bool IsLegal = (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0); 2378 LLVM_DEBUG(dbgs() << "Is imm " << C 2379 << " legal: " << (IsLegal ? "yes\n" : "no\n")); 2380 return IsLegal; 2381 } 2382 2383 // Can a (CMP op1, (sub 0, op2) be turned into a CMN instruction on 2384 // the grounds that "op1 - (-op2) == op1 + op2" ? Not always, the C and V flags 2385 // can be set differently by this operation. It comes down to whether 2386 // "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are then 2387 // everything is fine. If not then the optimization is wrong. Thus general 2388 // comparisons are only valid if op2 != 0. 2389 // 2390 // So, finally, the only LLVM-native comparisons that don't mention C and V 2391 // are SETEQ and SETNE. They're the only ones we can safely use CMN for in 2392 // the absence of information about op2. 2393 static bool isCMN(SDValue Op, ISD::CondCode CC) { 2394 return Op.getOpcode() == ISD::SUB && isNullConstant(Op.getOperand(0)) && 2395 (CC == ISD::SETEQ || CC == ISD::SETNE); 2396 } 2397 2398 static SDValue emitStrictFPComparison(SDValue LHS, SDValue RHS, const SDLoc &dl, 2399 SelectionDAG &DAG, SDValue Chain, 2400 bool IsSignaling) { 2401 EVT VT = LHS.getValueType(); 2402 assert(VT != MVT::f128); 2403 assert(VT != MVT::f16 && "Lowering of strict fp16 not yet implemented"); 2404 unsigned Opcode = 2405 IsSignaling ? AArch64ISD::STRICT_FCMPE : AArch64ISD::STRICT_FCMP; 2406 return DAG.getNode(Opcode, dl, {VT, MVT::Other}, {Chain, LHS, RHS}); 2407 } 2408 2409 static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC, 2410 const SDLoc &dl, SelectionDAG &DAG) { 2411 EVT VT = LHS.getValueType(); 2412 const bool FullFP16 = 2413 static_cast<const AArch64Subtarget &>(DAG.getSubtarget()).hasFullFP16(); 2414 2415 if (VT.isFloatingPoint()) { 2416 assert(VT != MVT::f128); 2417 if (VT == MVT::f16 && !FullFP16) { 2418 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, LHS); 2419 RHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, RHS); 2420 VT = MVT::f32; 2421 } 2422 return DAG.getNode(AArch64ISD::FCMP, dl, VT, LHS, RHS); 2423 } 2424 2425 // The CMP instruction is just an alias for SUBS, and representing it as 2426 // SUBS means that it's possible to get CSE with subtract operations. 2427 // A later phase can perform the optimization of setting the destination 2428 // register to WZR/XZR if it ends up being unused. 2429 unsigned Opcode = AArch64ISD::SUBS; 2430 2431 if (isCMN(RHS, CC)) { 2432 // Can we combine a (CMP op1, (sub 0, op2) into a CMN instruction ? 2433 Opcode = AArch64ISD::ADDS; 2434 RHS = RHS.getOperand(1); 2435 } else if (isCMN(LHS, CC)) { 2436 // As we are looking for EQ/NE compares, the operands can be commuted ; can 2437 // we combine a (CMP (sub 0, op1), op2) into a CMN instruction ? 2438 Opcode = AArch64ISD::ADDS; 2439 LHS = LHS.getOperand(1); 2440 } else if (isNullConstant(RHS) && !isUnsignedIntSetCC(CC)) { 2441 if (LHS.getOpcode() == ISD::AND) { 2442 // Similarly, (CMP (and X, Y), 0) can be implemented with a TST 2443 // (a.k.a. ANDS) except that the flags are only guaranteed to work for one 2444 // of the signed comparisons. 2445 const SDValue ANDSNode = DAG.getNode(AArch64ISD::ANDS, dl, 2446 DAG.getVTList(VT, MVT_CC), 2447 LHS.getOperand(0), 2448 LHS.getOperand(1)); 2449 // Replace all users of (and X, Y) with newly generated (ands X, Y) 2450 DAG.ReplaceAllUsesWith(LHS, ANDSNode); 2451 return ANDSNode.getValue(1); 2452 } else if (LHS.getOpcode() == AArch64ISD::ANDS) { 2453 // Use result of ANDS 2454 return LHS.getValue(1); 2455 } 2456 } 2457 2458 return DAG.getNode(Opcode, dl, DAG.getVTList(VT, MVT_CC), LHS, RHS) 2459 .getValue(1); 2460 } 2461 2462 /// \defgroup AArch64CCMP CMP;CCMP matching 2463 /// 2464 /// These functions deal with the formation of CMP;CCMP;... sequences. 2465 /// The CCMP/CCMN/FCCMP/FCCMPE instructions allow the conditional execution of 2466 /// a comparison. They set the NZCV flags to a predefined value if their 2467 /// predicate is false. This allows to express arbitrary conjunctions, for 2468 /// example "cmp 0 (and (setCA (cmp A)) (setCB (cmp B)))" 2469 /// expressed as: 2470 /// cmp A 2471 /// ccmp B, inv(CB), CA 2472 /// check for CB flags 2473 /// 2474 /// This naturally lets us implement chains of AND operations with SETCC 2475 /// operands. And we can even implement some other situations by transforming 2476 /// them: 2477 /// - We can implement (NEG SETCC) i.e. negating a single comparison by 2478 /// negating the flags used in a CCMP/FCCMP operations. 2479 /// - We can negate the result of a whole chain of CMP/CCMP/FCCMP operations 2480 /// by negating the flags we test for afterwards. i.e. 2481 /// NEG (CMP CCMP CCCMP ...) can be implemented. 2482 /// - Note that we can only ever negate all previously processed results. 2483 /// What we can not implement by flipping the flags to test is a negation 2484 /// of two sub-trees (because the negation affects all sub-trees emitted so 2485 /// far, so the 2nd sub-tree we emit would also affect the first). 2486 /// With those tools we can implement some OR operations: 2487 /// - (OR (SETCC A) (SETCC B)) can be implemented via: 2488 /// NEG (AND (NEG (SETCC A)) (NEG (SETCC B))) 2489 /// - After transforming OR to NEG/AND combinations we may be able to use NEG 2490 /// elimination rules from earlier to implement the whole thing as a 2491 /// CCMP/FCCMP chain. 2492 /// 2493 /// As complete example: 2494 /// or (or (setCA (cmp A)) (setCB (cmp B))) 2495 /// (and (setCC (cmp C)) (setCD (cmp D)))" 2496 /// can be reassociated to: 2497 /// or (and (setCC (cmp C)) setCD (cmp D)) 2498 // (or (setCA (cmp A)) (setCB (cmp B))) 2499 /// can be transformed to: 2500 /// not (and (not (and (setCC (cmp C)) (setCD (cmp D)))) 2501 /// (and (not (setCA (cmp A)) (not (setCB (cmp B))))))" 2502 /// which can be implemented as: 2503 /// cmp C 2504 /// ccmp D, inv(CD), CC 2505 /// ccmp A, CA, inv(CD) 2506 /// ccmp B, CB, inv(CA) 2507 /// check for CB flags 2508 /// 2509 /// A counterexample is "or (and A B) (and C D)" which translates to 2510 /// not (and (not (and (not A) (not B))) (not (and (not C) (not D)))), we 2511 /// can only implement 1 of the inner (not) operations, but not both! 2512 /// @{ 2513 2514 /// Create a conditional comparison; Use CCMP, CCMN or FCCMP as appropriate. 2515 static SDValue emitConditionalComparison(SDValue LHS, SDValue RHS, 2516 ISD::CondCode CC, SDValue CCOp, 2517 AArch64CC::CondCode Predicate, 2518 AArch64CC::CondCode OutCC, 2519 const SDLoc &DL, SelectionDAG &DAG) { 2520 unsigned Opcode = 0; 2521 const bool FullFP16 = 2522 static_cast<const AArch64Subtarget &>(DAG.getSubtarget()).hasFullFP16(); 2523 2524 if (LHS.getValueType().isFloatingPoint()) { 2525 assert(LHS.getValueType() != MVT::f128); 2526 if (LHS.getValueType() == MVT::f16 && !FullFP16) { 2527 LHS = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, LHS); 2528 RHS = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, RHS); 2529 } 2530 Opcode = AArch64ISD::FCCMP; 2531 } else if (RHS.getOpcode() == ISD::SUB) { 2532 SDValue SubOp0 = RHS.getOperand(0); 2533 if (isNullConstant(SubOp0) && (CC == ISD::SETEQ || CC == ISD::SETNE)) { 2534 // See emitComparison() on why we can only do this for SETEQ and SETNE. 2535 Opcode = AArch64ISD::CCMN; 2536 RHS = RHS.getOperand(1); 2537 } 2538 } 2539 if (Opcode == 0) 2540 Opcode = AArch64ISD::CCMP; 2541 2542 SDValue Condition = DAG.getConstant(Predicate, DL, MVT_CC); 2543 AArch64CC::CondCode InvOutCC = AArch64CC::getInvertedCondCode(OutCC); 2544 unsigned NZCV = AArch64CC::getNZCVToSatisfyCondCode(InvOutCC); 2545 SDValue NZCVOp = DAG.getConstant(NZCV, DL, MVT::i32); 2546 return DAG.getNode(Opcode, DL, MVT_CC, LHS, RHS, NZCVOp, Condition, CCOp); 2547 } 2548 2549 /// Returns true if @p Val is a tree of AND/OR/SETCC operations that can be 2550 /// expressed as a conjunction. See \ref AArch64CCMP. 2551 /// \param CanNegate Set to true if we can negate the whole sub-tree just by 2552 /// changing the conditions on the SETCC tests. 2553 /// (this means we can call emitConjunctionRec() with 2554 /// Negate==true on this sub-tree) 2555 /// \param MustBeFirst Set to true if this subtree needs to be negated and we 2556 /// cannot do the negation naturally. We are required to 2557 /// emit the subtree first in this case. 2558 /// \param WillNegate Is true if are called when the result of this 2559 /// subexpression must be negated. This happens when the 2560 /// outer expression is an OR. We can use this fact to know 2561 /// that we have a double negation (or (or ...) ...) that 2562 /// can be implemented for free. 2563 static bool canEmitConjunction(const SDValue Val, bool &CanNegate, 2564 bool &MustBeFirst, bool WillNegate, 2565 unsigned Depth = 0) { 2566 if (!Val.hasOneUse()) 2567 return false; 2568 unsigned Opcode = Val->getOpcode(); 2569 if (Opcode == ISD::SETCC) { 2570 if (Val->getOperand(0).getValueType() == MVT::f128) 2571 return false; 2572 CanNegate = true; 2573 MustBeFirst = false; 2574 return true; 2575 } 2576 // Protect against exponential runtime and stack overflow. 2577 if (Depth > 6) 2578 return false; 2579 if (Opcode == ISD::AND || Opcode == ISD::OR) { 2580 bool IsOR = Opcode == ISD::OR; 2581 SDValue O0 = Val->getOperand(0); 2582 SDValue O1 = Val->getOperand(1); 2583 bool CanNegateL; 2584 bool MustBeFirstL; 2585 if (!canEmitConjunction(O0, CanNegateL, MustBeFirstL, IsOR, Depth+1)) 2586 return false; 2587 bool CanNegateR; 2588 bool MustBeFirstR; 2589 if (!canEmitConjunction(O1, CanNegateR, MustBeFirstR, IsOR, Depth+1)) 2590 return false; 2591 2592 if (MustBeFirstL && MustBeFirstR) 2593 return false; 2594 2595 if (IsOR) { 2596 // For an OR expression we need to be able to naturally negate at least 2597 // one side or we cannot do the transformation at all. 2598 if (!CanNegateL && !CanNegateR) 2599 return false; 2600 // If we the result of the OR will be negated and we can naturally negate 2601 // the leafs, then this sub-tree as a whole negates naturally. 2602 CanNegate = WillNegate && CanNegateL && CanNegateR; 2603 // If we cannot naturally negate the whole sub-tree, then this must be 2604 // emitted first. 2605 MustBeFirst = !CanNegate; 2606 } else { 2607 assert(Opcode == ISD::AND && "Must be OR or AND"); 2608 // We cannot naturally negate an AND operation. 2609 CanNegate = false; 2610 MustBeFirst = MustBeFirstL || MustBeFirstR; 2611 } 2612 return true; 2613 } 2614 return false; 2615 } 2616 2617 /// Emit conjunction or disjunction tree with the CMP/FCMP followed by a chain 2618 /// of CCMP/CFCMP ops. See @ref AArch64CCMP. 2619 /// Tries to transform the given i1 producing node @p Val to a series compare 2620 /// and conditional compare operations. @returns an NZCV flags producing node 2621 /// and sets @p OutCC to the flags that should be tested or returns SDValue() if 2622 /// transformation was not possible. 2623 /// \p Negate is true if we want this sub-tree being negated just by changing 2624 /// SETCC conditions. 2625 static SDValue emitConjunctionRec(SelectionDAG &DAG, SDValue Val, 2626 AArch64CC::CondCode &OutCC, bool Negate, SDValue CCOp, 2627 AArch64CC::CondCode Predicate) { 2628 // We're at a tree leaf, produce a conditional comparison operation. 2629 unsigned Opcode = Val->getOpcode(); 2630 if (Opcode == ISD::SETCC) { 2631 SDValue LHS = Val->getOperand(0); 2632 SDValue RHS = Val->getOperand(1); 2633 ISD::CondCode CC = cast<CondCodeSDNode>(Val->getOperand(2))->get(); 2634 bool isInteger = LHS.getValueType().isInteger(); 2635 if (Negate) 2636 CC = getSetCCInverse(CC, LHS.getValueType()); 2637 SDLoc DL(Val); 2638 // Determine OutCC and handle FP special case. 2639 if (isInteger) { 2640 OutCC = changeIntCCToAArch64CC(CC); 2641 } else { 2642 assert(LHS.getValueType().isFloatingPoint()); 2643 AArch64CC::CondCode ExtraCC; 2644 changeFPCCToANDAArch64CC(CC, OutCC, ExtraCC); 2645 // Some floating point conditions can't be tested with a single condition 2646 // code. Construct an additional comparison in this case. 2647 if (ExtraCC != AArch64CC::AL) { 2648 SDValue ExtraCmp; 2649 if (!CCOp.getNode()) 2650 ExtraCmp = emitComparison(LHS, RHS, CC, DL, DAG); 2651 else 2652 ExtraCmp = emitConditionalComparison(LHS, RHS, CC, CCOp, Predicate, 2653 ExtraCC, DL, DAG); 2654 CCOp = ExtraCmp; 2655 Predicate = ExtraCC; 2656 } 2657 } 2658 2659 // Produce a normal comparison if we are first in the chain 2660 if (!CCOp) 2661 return emitComparison(LHS, RHS, CC, DL, DAG); 2662 // Otherwise produce a ccmp. 2663 return emitConditionalComparison(LHS, RHS, CC, CCOp, Predicate, OutCC, DL, 2664 DAG); 2665 } 2666 assert(Val->hasOneUse() && "Valid conjunction/disjunction tree"); 2667 2668 bool IsOR = Opcode == ISD::OR; 2669 2670 SDValue LHS = Val->getOperand(0); 2671 bool CanNegateL; 2672 bool MustBeFirstL; 2673 bool ValidL = canEmitConjunction(LHS, CanNegateL, MustBeFirstL, IsOR); 2674 assert(ValidL && "Valid conjunction/disjunction tree"); 2675 (void)ValidL; 2676 2677 SDValue RHS = Val->getOperand(1); 2678 bool CanNegateR; 2679 bool MustBeFirstR; 2680 bool ValidR = canEmitConjunction(RHS, CanNegateR, MustBeFirstR, IsOR); 2681 assert(ValidR && "Valid conjunction/disjunction tree"); 2682 (void)ValidR; 2683 2684 // Swap sub-tree that must come first to the right side. 2685 if (MustBeFirstL) { 2686 assert(!MustBeFirstR && "Valid conjunction/disjunction tree"); 2687 std::swap(LHS, RHS); 2688 std::swap(CanNegateL, CanNegateR); 2689 std::swap(MustBeFirstL, MustBeFirstR); 2690 } 2691 2692 bool NegateR; 2693 bool NegateAfterR; 2694 bool NegateL; 2695 bool NegateAfterAll; 2696 if (Opcode == ISD::OR) { 2697 // Swap the sub-tree that we can negate naturally to the left. 2698 if (!CanNegateL) { 2699 assert(CanNegateR && "at least one side must be negatable"); 2700 assert(!MustBeFirstR && "invalid conjunction/disjunction tree"); 2701 assert(!Negate); 2702 std::swap(LHS, RHS); 2703 NegateR = false; 2704 NegateAfterR = true; 2705 } else { 2706 // Negate the left sub-tree if possible, otherwise negate the result. 2707 NegateR = CanNegateR; 2708 NegateAfterR = !CanNegateR; 2709 } 2710 NegateL = true; 2711 NegateAfterAll = !Negate; 2712 } else { 2713 assert(Opcode == ISD::AND && "Valid conjunction/disjunction tree"); 2714 assert(!Negate && "Valid conjunction/disjunction tree"); 2715 2716 NegateL = false; 2717 NegateR = false; 2718 NegateAfterR = false; 2719 NegateAfterAll = false; 2720 } 2721 2722 // Emit sub-trees. 2723 AArch64CC::CondCode RHSCC; 2724 SDValue CmpR = emitConjunctionRec(DAG, RHS, RHSCC, NegateR, CCOp, Predicate); 2725 if (NegateAfterR) 2726 RHSCC = AArch64CC::getInvertedCondCode(RHSCC); 2727 SDValue CmpL = emitConjunctionRec(DAG, LHS, OutCC, NegateL, CmpR, RHSCC); 2728 if (NegateAfterAll) 2729 OutCC = AArch64CC::getInvertedCondCode(OutCC); 2730 return CmpL; 2731 } 2732 2733 /// Emit expression as a conjunction (a series of CCMP/CFCMP ops). 2734 /// In some cases this is even possible with OR operations in the expression. 2735 /// See \ref AArch64CCMP. 2736 /// \see emitConjunctionRec(). 2737 static SDValue emitConjunction(SelectionDAG &DAG, SDValue Val, 2738 AArch64CC::CondCode &OutCC) { 2739 bool DummyCanNegate; 2740 bool DummyMustBeFirst; 2741 if (!canEmitConjunction(Val, DummyCanNegate, DummyMustBeFirst, false)) 2742 return SDValue(); 2743 2744 return emitConjunctionRec(DAG, Val, OutCC, false, SDValue(), AArch64CC::AL); 2745 } 2746 2747 /// @} 2748 2749 /// Returns how profitable it is to fold a comparison's operand's shift and/or 2750 /// extension operations. 2751 static unsigned getCmpOperandFoldingProfit(SDValue Op) { 2752 auto isSupportedExtend = [&](SDValue V) { 2753 if (V.getOpcode() == ISD::SIGN_EXTEND_INREG) 2754 return true; 2755 2756 if (V.getOpcode() == ISD::AND) 2757 if (ConstantSDNode *MaskCst = dyn_cast<ConstantSDNode>(V.getOperand(1))) { 2758 uint64_t Mask = MaskCst->getZExtValue(); 2759 return (Mask == 0xFF || Mask == 0xFFFF || Mask == 0xFFFFFFFF); 2760 } 2761 2762 return false; 2763 }; 2764 2765 if (!Op.hasOneUse()) 2766 return 0; 2767 2768 if (isSupportedExtend(Op)) 2769 return 1; 2770 2771 unsigned Opc = Op.getOpcode(); 2772 if (Opc == ISD::SHL || Opc == ISD::SRL || Opc == ISD::SRA) 2773 if (ConstantSDNode *ShiftCst = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 2774 uint64_t Shift = ShiftCst->getZExtValue(); 2775 if (isSupportedExtend(Op.getOperand(0))) 2776 return (Shift <= 4) ? 2 : 1; 2777 EVT VT = Op.getValueType(); 2778 if ((VT == MVT::i32 && Shift <= 31) || (VT == MVT::i64 && Shift <= 63)) 2779 return 1; 2780 } 2781 2782 return 0; 2783 } 2784 2785 static SDValue getAArch64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC, 2786 SDValue &AArch64cc, SelectionDAG &DAG, 2787 const SDLoc &dl) { 2788 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) { 2789 EVT VT = RHS.getValueType(); 2790 uint64_t C = RHSC->getZExtValue(); 2791 if (!isLegalArithImmed(C)) { 2792 // Constant does not fit, try adjusting it by one? 2793 switch (CC) { 2794 default: 2795 break; 2796 case ISD::SETLT: 2797 case ISD::SETGE: 2798 if ((VT == MVT::i32 && C != 0x80000000 && 2799 isLegalArithImmed((uint32_t)(C - 1))) || 2800 (VT == MVT::i64 && C != 0x80000000ULL && 2801 isLegalArithImmed(C - 1ULL))) { 2802 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT; 2803 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1; 2804 RHS = DAG.getConstant(C, dl, VT); 2805 } 2806 break; 2807 case ISD::SETULT: 2808 case ISD::SETUGE: 2809 if ((VT == MVT::i32 && C != 0 && 2810 isLegalArithImmed((uint32_t)(C - 1))) || 2811 (VT == MVT::i64 && C != 0ULL && isLegalArithImmed(C - 1ULL))) { 2812 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT; 2813 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1; 2814 RHS = DAG.getConstant(C, dl, VT); 2815 } 2816 break; 2817 case ISD::SETLE: 2818 case ISD::SETGT: 2819 if ((VT == MVT::i32 && C != INT32_MAX && 2820 isLegalArithImmed((uint32_t)(C + 1))) || 2821 (VT == MVT::i64 && C != INT64_MAX && 2822 isLegalArithImmed(C + 1ULL))) { 2823 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE; 2824 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1; 2825 RHS = DAG.getConstant(C, dl, VT); 2826 } 2827 break; 2828 case ISD::SETULE: 2829 case ISD::SETUGT: 2830 if ((VT == MVT::i32 && C != UINT32_MAX && 2831 isLegalArithImmed((uint32_t)(C + 1))) || 2832 (VT == MVT::i64 && C != UINT64_MAX && 2833 isLegalArithImmed(C + 1ULL))) { 2834 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE; 2835 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1; 2836 RHS = DAG.getConstant(C, dl, VT); 2837 } 2838 break; 2839 } 2840 } 2841 } 2842 2843 // Comparisons are canonicalized so that the RHS operand is simpler than the 2844 // LHS one, the extreme case being when RHS is an immediate. However, AArch64 2845 // can fold some shift+extend operations on the RHS operand, so swap the 2846 // operands if that can be done. 2847 // 2848 // For example: 2849 // lsl w13, w11, #1 2850 // cmp w13, w12 2851 // can be turned into: 2852 // cmp w12, w11, lsl #1 2853 if (!isa<ConstantSDNode>(RHS) || 2854 !isLegalArithImmed(cast<ConstantSDNode>(RHS)->getZExtValue())) { 2855 SDValue TheLHS = isCMN(LHS, CC) ? LHS.getOperand(1) : LHS; 2856 2857 if (getCmpOperandFoldingProfit(TheLHS) > getCmpOperandFoldingProfit(RHS)) { 2858 std::swap(LHS, RHS); 2859 CC = ISD::getSetCCSwappedOperands(CC); 2860 } 2861 } 2862 2863 SDValue Cmp; 2864 AArch64CC::CondCode AArch64CC; 2865 if ((CC == ISD::SETEQ || CC == ISD::SETNE) && isa<ConstantSDNode>(RHS)) { 2866 const ConstantSDNode *RHSC = cast<ConstantSDNode>(RHS); 2867 2868 // The imm operand of ADDS is an unsigned immediate, in the range 0 to 4095. 2869 // For the i8 operand, the largest immediate is 255, so this can be easily 2870 // encoded in the compare instruction. For the i16 operand, however, the 2871 // largest immediate cannot be encoded in the compare. 2872 // Therefore, use a sign extending load and cmn to avoid materializing the 2873 // -1 constant. For example, 2874 // movz w1, #65535 2875 // ldrh w0, [x0, #0] 2876 // cmp w0, w1 2877 // > 2878 // ldrsh w0, [x0, #0] 2879 // cmn w0, #1 2880 // Fundamental, we're relying on the property that (zext LHS) == (zext RHS) 2881 // if and only if (sext LHS) == (sext RHS). The checks are in place to 2882 // ensure both the LHS and RHS are truly zero extended and to make sure the 2883 // transformation is profitable. 2884 if ((RHSC->getZExtValue() >> 16 == 0) && isa<LoadSDNode>(LHS) && 2885 cast<LoadSDNode>(LHS)->getExtensionType() == ISD::ZEXTLOAD && 2886 cast<LoadSDNode>(LHS)->getMemoryVT() == MVT::i16 && 2887 LHS.getNode()->hasNUsesOfValue(1, 0)) { 2888 int16_t ValueofRHS = cast<ConstantSDNode>(RHS)->getZExtValue(); 2889 if (ValueofRHS < 0 && isLegalArithImmed(-ValueofRHS)) { 2890 SDValue SExt = 2891 DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, LHS.getValueType(), LHS, 2892 DAG.getValueType(MVT::i16)); 2893 Cmp = emitComparison(SExt, DAG.getConstant(ValueofRHS, dl, 2894 RHS.getValueType()), 2895 CC, dl, DAG); 2896 AArch64CC = changeIntCCToAArch64CC(CC); 2897 } 2898 } 2899 2900 if (!Cmp && (RHSC->isNullValue() || RHSC->isOne())) { 2901 if ((Cmp = emitConjunction(DAG, LHS, AArch64CC))) { 2902 if ((CC == ISD::SETNE) ^ RHSC->isNullValue()) 2903 AArch64CC = AArch64CC::getInvertedCondCode(AArch64CC); 2904 } 2905 } 2906 } 2907 2908 if (!Cmp) { 2909 Cmp = emitComparison(LHS, RHS, CC, dl, DAG); 2910 AArch64CC = changeIntCCToAArch64CC(CC); 2911 } 2912 AArch64cc = DAG.getConstant(AArch64CC, dl, MVT_CC); 2913 return Cmp; 2914 } 2915 2916 static std::pair<SDValue, SDValue> 2917 getAArch64XALUOOp(AArch64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) { 2918 assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) && 2919 "Unsupported value type"); 2920 SDValue Value, Overflow; 2921 SDLoc DL(Op); 2922 SDValue LHS = Op.getOperand(0); 2923 SDValue RHS = Op.getOperand(1); 2924 unsigned Opc = 0; 2925 switch (Op.getOpcode()) { 2926 default: 2927 llvm_unreachable("Unknown overflow instruction!"); 2928 case ISD::SADDO: 2929 Opc = AArch64ISD::ADDS; 2930 CC = AArch64CC::VS; 2931 break; 2932 case ISD::UADDO: 2933 Opc = AArch64ISD::ADDS; 2934 CC = AArch64CC::HS; 2935 break; 2936 case ISD::SSUBO: 2937 Opc = AArch64ISD::SUBS; 2938 CC = AArch64CC::VS; 2939 break; 2940 case ISD::USUBO: 2941 Opc = AArch64ISD::SUBS; 2942 CC = AArch64CC::LO; 2943 break; 2944 // Multiply needs a little bit extra work. 2945 case ISD::SMULO: 2946 case ISD::UMULO: { 2947 CC = AArch64CC::NE; 2948 bool IsSigned = Op.getOpcode() == ISD::SMULO; 2949 if (Op.getValueType() == MVT::i32) { 2950 unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; 2951 // For a 32 bit multiply with overflow check we want the instruction 2952 // selector to generate a widening multiply (SMADDL/UMADDL). For that we 2953 // need to generate the following pattern: 2954 // (i64 add 0, (i64 mul (i64 sext|zext i32 %a), (i64 sext|zext i32 %b)) 2955 LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS); 2956 RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS); 2957 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS); 2958 SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Mul, 2959 DAG.getConstant(0, DL, MVT::i64)); 2960 // On AArch64 the upper 32 bits are always zero extended for a 32 bit 2961 // operation. We need to clear out the upper 32 bits, because we used a 2962 // widening multiply that wrote all 64 bits. In the end this should be a 2963 // noop. 2964 Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Add); 2965 if (IsSigned) { 2966 // The signed overflow check requires more than just a simple check for 2967 // any bit set in the upper 32 bits of the result. These bits could be 2968 // just the sign bits of a negative number. To perform the overflow 2969 // check we have to arithmetic shift right the 32nd bit of the result by 2970 // 31 bits. Then we compare the result to the upper 32 bits. 2971 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Add, 2972 DAG.getConstant(32, DL, MVT::i64)); 2973 UpperBits = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, UpperBits); 2974 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i32, Value, 2975 DAG.getConstant(31, DL, MVT::i64)); 2976 // It is important that LowerBits is last, otherwise the arithmetic 2977 // shift will not be folded into the compare (SUBS). 2978 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32); 2979 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits) 2980 .getValue(1); 2981 } else { 2982 // The overflow check for unsigned multiply is easy. We only need to 2983 // check if any of the upper 32 bits are set. This can be done with a 2984 // CMP (shifted register). For that we need to generate the following 2985 // pattern: 2986 // (i64 AArch64ISD::SUBS i64 0, (i64 srl i64 %Mul, i64 32) 2987 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul, 2988 DAG.getConstant(32, DL, MVT::i64)); 2989 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32); 2990 Overflow = 2991 DAG.getNode(AArch64ISD::SUBS, DL, VTs, 2992 DAG.getConstant(0, DL, MVT::i64), 2993 UpperBits).getValue(1); 2994 } 2995 break; 2996 } 2997 assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type"); 2998 // For the 64 bit multiply 2999 Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS); 3000 if (IsSigned) { 3001 SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS); 3002 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value, 3003 DAG.getConstant(63, DL, MVT::i64)); 3004 // It is important that LowerBits is last, otherwise the arithmetic 3005 // shift will not be folded into the compare (SUBS). 3006 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32); 3007 Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits) 3008 .getValue(1); 3009 } else { 3010 SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS); 3011 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32); 3012 Overflow = 3013 DAG.getNode(AArch64ISD::SUBS, DL, VTs, 3014 DAG.getConstant(0, DL, MVT::i64), 3015 UpperBits).getValue(1); 3016 } 3017 break; 3018 } 3019 } // switch (...) 3020 3021 if (Opc) { 3022 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32); 3023 3024 // Emit the AArch64 operation with overflow check. 3025 Value = DAG.getNode(Opc, DL, VTs, LHS, RHS); 3026 Overflow = Value.getValue(1); 3027 } 3028 return std::make_pair(Value, Overflow); 3029 } 3030 3031 SDValue AArch64TargetLowering::LowerXOR(SDValue Op, SelectionDAG &DAG) const { 3032 if (useSVEForFixedLengthVectorVT(Op.getValueType())) 3033 return LowerToScalableOp(Op, DAG); 3034 3035 SDValue Sel = Op.getOperand(0); 3036 SDValue Other = Op.getOperand(1); 3037 SDLoc dl(Sel); 3038 3039 // If the operand is an overflow checking operation, invert the condition 3040 // code and kill the Not operation. I.e., transform: 3041 // (xor (overflow_op_bool, 1)) 3042 // --> 3043 // (csel 1, 0, invert(cc), overflow_op_bool) 3044 // ... which later gets transformed to just a cset instruction with an 3045 // inverted condition code, rather than a cset + eor sequence. 3046 if (isOneConstant(Other) && ISD::isOverflowIntrOpRes(Sel)) { 3047 // Only lower legal XALUO ops. 3048 if (!DAG.getTargetLoweringInfo().isTypeLegal(Sel->getValueType(0))) 3049 return SDValue(); 3050 3051 SDValue TVal = DAG.getConstant(1, dl, MVT::i32); 3052 SDValue FVal = DAG.getConstant(0, dl, MVT::i32); 3053 AArch64CC::CondCode CC; 3054 SDValue Value, Overflow; 3055 std::tie(Value, Overflow) = getAArch64XALUOOp(CC, Sel.getValue(0), DAG); 3056 SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), dl, MVT::i32); 3057 return DAG.getNode(AArch64ISD::CSEL, dl, Op.getValueType(), TVal, FVal, 3058 CCVal, Overflow); 3059 } 3060 // If neither operand is a SELECT_CC, give up. 3061 if (Sel.getOpcode() != ISD::SELECT_CC) 3062 std::swap(Sel, Other); 3063 if (Sel.getOpcode() != ISD::SELECT_CC) 3064 return Op; 3065 3066 // The folding we want to perform is: 3067 // (xor x, (select_cc a, b, cc, 0, -1) ) 3068 // --> 3069 // (csel x, (xor x, -1), cc ...) 3070 // 3071 // The latter will get matched to a CSINV instruction. 3072 3073 ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get(); 3074 SDValue LHS = Sel.getOperand(0); 3075 SDValue RHS = Sel.getOperand(1); 3076 SDValue TVal = Sel.getOperand(2); 3077 SDValue FVal = Sel.getOperand(3); 3078 3079 // FIXME: This could be generalized to non-integer comparisons. 3080 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64) 3081 return Op; 3082 3083 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal); 3084 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal); 3085 3086 // The values aren't constants, this isn't the pattern we're looking for. 3087 if (!CFVal || !CTVal) 3088 return Op; 3089 3090 // We can commute the SELECT_CC by inverting the condition. This 3091 // might be needed to make this fit into a CSINV pattern. 3092 if (CTVal->isAllOnesValue() && CFVal->isNullValue()) { 3093 std::swap(TVal, FVal); 3094 std::swap(CTVal, CFVal); 3095 CC = ISD::getSetCCInverse(CC, LHS.getValueType()); 3096 } 3097 3098 // If the constants line up, perform the transform! 3099 if (CTVal->isNullValue() && CFVal->isAllOnesValue()) { 3100 SDValue CCVal; 3101 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl); 3102 3103 FVal = Other; 3104 TVal = DAG.getNode(ISD::XOR, dl, Other.getValueType(), Other, 3105 DAG.getConstant(-1ULL, dl, Other.getValueType())); 3106 3107 return DAG.getNode(AArch64ISD::CSEL, dl, Sel.getValueType(), FVal, TVal, 3108 CCVal, Cmp); 3109 } 3110 3111 return Op; 3112 } 3113 3114 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) { 3115 EVT VT = Op.getValueType(); 3116 3117 // Let legalize expand this if it isn't a legal type yet. 3118 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT)) 3119 return SDValue(); 3120 3121 SDVTList VTs = DAG.getVTList(VT, MVT::i32); 3122 3123 unsigned Opc; 3124 bool ExtraOp = false; 3125 switch (Op.getOpcode()) { 3126 default: 3127 llvm_unreachable("Invalid code"); 3128 case ISD::ADDC: 3129 Opc = AArch64ISD::ADDS; 3130 break; 3131 case ISD::SUBC: 3132 Opc = AArch64ISD::SUBS; 3133 break; 3134 case ISD::ADDE: 3135 Opc = AArch64ISD::ADCS; 3136 ExtraOp = true; 3137 break; 3138 case ISD::SUBE: 3139 Opc = AArch64ISD::SBCS; 3140 ExtraOp = true; 3141 break; 3142 } 3143 3144 if (!ExtraOp) 3145 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1)); 3146 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1), 3147 Op.getOperand(2)); 3148 } 3149 3150 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) { 3151 // Let legalize expand this if it isn't a legal type yet. 3152 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType())) 3153 return SDValue(); 3154 3155 SDLoc dl(Op); 3156 AArch64CC::CondCode CC; 3157 // The actual operation that sets the overflow or carry flag. 3158 SDValue Value, Overflow; 3159 std::tie(Value, Overflow) = getAArch64XALUOOp(CC, Op, DAG); 3160 3161 // We use 0 and 1 as false and true values. 3162 SDValue TVal = DAG.getConstant(1, dl, MVT::i32); 3163 SDValue FVal = DAG.getConstant(0, dl, MVT::i32); 3164 3165 // We use an inverted condition, because the conditional select is inverted 3166 // too. This will allow it to be selected to a single instruction: 3167 // CSINC Wd, WZR, WZR, invert(cond). 3168 SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), dl, MVT::i32); 3169 Overflow = DAG.getNode(AArch64ISD::CSEL, dl, MVT::i32, FVal, TVal, 3170 CCVal, Overflow); 3171 3172 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32); 3173 return DAG.getNode(ISD::MERGE_VALUES, dl, VTs, Value, Overflow); 3174 } 3175 3176 // Prefetch operands are: 3177 // 1: Address to prefetch 3178 // 2: bool isWrite 3179 // 3: int locality (0 = no locality ... 3 = extreme locality) 3180 // 4: bool isDataCache 3181 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) { 3182 SDLoc DL(Op); 3183 unsigned IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue(); 3184 unsigned Locality = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue(); 3185 unsigned IsData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue(); 3186 3187 bool IsStream = !Locality; 3188 // When the locality number is set 3189 if (Locality) { 3190 // The front-end should have filtered out the out-of-range values 3191 assert(Locality <= 3 && "Prefetch locality out-of-range"); 3192 // The locality degree is the opposite of the cache speed. 3193 // Put the number the other way around. 3194 // The encoding starts at 0 for level 1 3195 Locality = 3 - Locality; 3196 } 3197 3198 // built the mask value encoding the expected behavior. 3199 unsigned PrfOp = (IsWrite << 4) | // Load/Store bit 3200 (!IsData << 3) | // IsDataCache bit 3201 (Locality << 1) | // Cache level bits 3202 (unsigned)IsStream; // Stream bit 3203 return DAG.getNode(AArch64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0), 3204 DAG.getConstant(PrfOp, DL, MVT::i32), Op.getOperand(1)); 3205 } 3206 3207 SDValue AArch64TargetLowering::LowerFP_EXTEND(SDValue Op, 3208 SelectionDAG &DAG) const { 3209 if (Op.getValueType().isScalableVector()) 3210 return LowerToPredicatedOp(Op, DAG, AArch64ISD::FP_EXTEND_MERGE_PASSTHRU); 3211 3212 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering"); 3213 return SDValue(); 3214 } 3215 3216 SDValue AArch64TargetLowering::LowerFP_ROUND(SDValue Op, 3217 SelectionDAG &DAG) const { 3218 if (Op.getValueType().isScalableVector()) 3219 return LowerToPredicatedOp(Op, DAG, AArch64ISD::FP_ROUND_MERGE_PASSTHRU); 3220 3221 bool IsStrict = Op->isStrictFPOpcode(); 3222 SDValue SrcVal = Op.getOperand(IsStrict ? 1 : 0); 3223 EVT SrcVT = SrcVal.getValueType(); 3224 3225 if (SrcVT != MVT::f128) { 3226 // Expand cases where the input is a vector bigger than NEON. 3227 if (useSVEForFixedLengthVectorVT(SrcVT)) 3228 return SDValue(); 3229 3230 // It's legal except when f128 is involved 3231 return Op; 3232 } 3233 3234 return SDValue(); 3235 } 3236 3237 SDValue AArch64TargetLowering::LowerVectorFP_TO_INT(SDValue Op, 3238 SelectionDAG &DAG) const { 3239 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp. 3240 // Any additional optimization in this function should be recorded 3241 // in the cost tables. 3242 EVT InVT = Op.getOperand(0).getValueType(); 3243 EVT VT = Op.getValueType(); 3244 3245 if (VT.isScalableVector()) { 3246 unsigned Opcode = Op.getOpcode() == ISD::FP_TO_UINT 3247 ? AArch64ISD::FCVTZU_MERGE_PASSTHRU 3248 : AArch64ISD::FCVTZS_MERGE_PASSTHRU; 3249 return LowerToPredicatedOp(Op, DAG, Opcode); 3250 } 3251 3252 unsigned NumElts = InVT.getVectorNumElements(); 3253 3254 // f16 conversions are promoted to f32 when full fp16 is not supported. 3255 if (InVT.getVectorElementType() == MVT::f16 && 3256 !Subtarget->hasFullFP16()) { 3257 MVT NewVT = MVT::getVectorVT(MVT::f32, NumElts); 3258 SDLoc dl(Op); 3259 return DAG.getNode( 3260 Op.getOpcode(), dl, Op.getValueType(), 3261 DAG.getNode(ISD::FP_EXTEND, dl, NewVT, Op.getOperand(0))); 3262 } 3263 3264 uint64_t VTSize = VT.getFixedSizeInBits(); 3265 uint64_t InVTSize = InVT.getFixedSizeInBits(); 3266 if (VTSize < InVTSize) { 3267 SDLoc dl(Op); 3268 SDValue Cv = 3269 DAG.getNode(Op.getOpcode(), dl, InVT.changeVectorElementTypeToInteger(), 3270 Op.getOperand(0)); 3271 return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv); 3272 } 3273 3274 if (VTSize > InVTSize) { 3275 SDLoc dl(Op); 3276 MVT ExtVT = 3277 MVT::getVectorVT(MVT::getFloatingPointVT(VT.getScalarSizeInBits()), 3278 VT.getVectorNumElements()); 3279 SDValue Ext = DAG.getNode(ISD::FP_EXTEND, dl, ExtVT, Op.getOperand(0)); 3280 return DAG.getNode(Op.getOpcode(), dl, VT, Ext); 3281 } 3282 3283 // Type changing conversions are illegal. 3284 return Op; 3285 } 3286 3287 SDValue AArch64TargetLowering::LowerFP_TO_INT(SDValue Op, 3288 SelectionDAG &DAG) const { 3289 bool IsStrict = Op->isStrictFPOpcode(); 3290 SDValue SrcVal = Op.getOperand(IsStrict ? 1 : 0); 3291 3292 if (SrcVal.getValueType().isVector()) 3293 return LowerVectorFP_TO_INT(Op, DAG); 3294 3295 // f16 conversions are promoted to f32 when full fp16 is not supported. 3296 if (SrcVal.getValueType() == MVT::f16 && !Subtarget->hasFullFP16()) { 3297 assert(!IsStrict && "Lowering of strict fp16 not yet implemented"); 3298 SDLoc dl(Op); 3299 return DAG.getNode( 3300 Op.getOpcode(), dl, Op.getValueType(), 3301 DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, SrcVal)); 3302 } 3303 3304 if (SrcVal.getValueType() != MVT::f128) { 3305 // It's legal except when f128 is involved 3306 return Op; 3307 } 3308 3309 return SDValue(); 3310 } 3311 3312 SDValue AArch64TargetLowering::LowerFP_TO_INT_SAT(SDValue Op, 3313 SelectionDAG &DAG) const { 3314 // AArch64 FP-to-int conversions saturate to the destination register size, so 3315 // we can lower common saturating conversions to simple instructions. 3316 SDValue SrcVal = Op.getOperand(0); 3317 3318 EVT SrcVT = SrcVal.getValueType(); 3319 EVT DstVT = Op.getValueType(); 3320 3321 EVT SatVT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 3322 uint64_t SatWidth = SatVT.getScalarSizeInBits(); 3323 uint64_t DstWidth = DstVT.getScalarSizeInBits(); 3324 assert(SatWidth <= DstWidth && "Saturation width cannot exceed result width"); 3325 3326 // TODO: Support lowering of NEON and SVE conversions. 3327 if (SrcVT.isVector()) 3328 return SDValue(); 3329 3330 // TODO: Saturate to SatWidth explicitly. 3331 if (SatWidth != DstWidth) 3332 return SDValue(); 3333 3334 // In the absence of FP16 support, promote f32 to f16, like LowerFP_TO_INT(). 3335 if (SrcVT == MVT::f16 && !Subtarget->hasFullFP16()) 3336 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(), 3337 DAG.getNode(ISD::FP_EXTEND, SDLoc(Op), MVT::f32, SrcVal), 3338 Op.getOperand(1)); 3339 3340 // Cases that we can emit directly. 3341 if ((SrcVT == MVT::f64 || SrcVT == MVT::f32 || 3342 (SrcVT == MVT::f16 && Subtarget->hasFullFP16())) && 3343 (DstVT == MVT::i64 || DstVT == MVT::i32)) 3344 return Op; 3345 3346 // For all other cases, fall back on the expanded form. 3347 return SDValue(); 3348 } 3349 3350 SDValue AArch64TargetLowering::LowerVectorINT_TO_FP(SDValue Op, 3351 SelectionDAG &DAG) const { 3352 // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp. 3353 // Any additional optimization in this function should be recorded 3354 // in the cost tables. 3355 EVT VT = Op.getValueType(); 3356 SDLoc dl(Op); 3357 SDValue In = Op.getOperand(0); 3358 EVT InVT = In.getValueType(); 3359 unsigned Opc = Op.getOpcode(); 3360 bool IsSigned = Opc == ISD::SINT_TO_FP || Opc == ISD::STRICT_SINT_TO_FP; 3361 3362 if (VT.isScalableVector()) { 3363 if (InVT.getVectorElementType() == MVT::i1) { 3364 // We can't directly extend an SVE predicate; extend it first. 3365 unsigned CastOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; 3366 EVT CastVT = getPromotedVTForPredicate(InVT); 3367 In = DAG.getNode(CastOpc, dl, CastVT, In); 3368 return DAG.getNode(Opc, dl, VT, In); 3369 } 3370 3371 unsigned Opcode = IsSigned ? AArch64ISD::SINT_TO_FP_MERGE_PASSTHRU 3372 : AArch64ISD::UINT_TO_FP_MERGE_PASSTHRU; 3373 return LowerToPredicatedOp(Op, DAG, Opcode); 3374 } 3375 3376 uint64_t VTSize = VT.getFixedSizeInBits(); 3377 uint64_t InVTSize = InVT.getFixedSizeInBits(); 3378 if (VTSize < InVTSize) { 3379 MVT CastVT = 3380 MVT::getVectorVT(MVT::getFloatingPointVT(InVT.getScalarSizeInBits()), 3381 InVT.getVectorNumElements()); 3382 In = DAG.getNode(Opc, dl, CastVT, In); 3383 return DAG.getNode(ISD::FP_ROUND, dl, VT, In, DAG.getIntPtrConstant(0, dl)); 3384 } 3385 3386 if (VTSize > InVTSize) { 3387 unsigned CastOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; 3388 EVT CastVT = VT.changeVectorElementTypeToInteger(); 3389 In = DAG.getNode(CastOpc, dl, CastVT, In); 3390 return DAG.getNode(Opc, dl, VT, In); 3391 } 3392 3393 return Op; 3394 } 3395 3396 SDValue AArch64TargetLowering::LowerINT_TO_FP(SDValue Op, 3397 SelectionDAG &DAG) const { 3398 if (Op.getValueType().isVector()) 3399 return LowerVectorINT_TO_FP(Op, DAG); 3400 3401 bool IsStrict = Op->isStrictFPOpcode(); 3402 SDValue SrcVal = Op.getOperand(IsStrict ? 1 : 0); 3403 3404 // f16 conversions are promoted to f32 when full fp16 is not supported. 3405 if (Op.getValueType() == MVT::f16 && 3406 !Subtarget->hasFullFP16()) { 3407 assert(!IsStrict && "Lowering of strict fp16 not yet implemented"); 3408 SDLoc dl(Op); 3409 return DAG.getNode( 3410 ISD::FP_ROUND, dl, MVT::f16, 3411 DAG.getNode(Op.getOpcode(), dl, MVT::f32, SrcVal), 3412 DAG.getIntPtrConstant(0, dl)); 3413 } 3414 3415 // i128 conversions are libcalls. 3416 if (SrcVal.getValueType() == MVT::i128) 3417 return SDValue(); 3418 3419 // Other conversions are legal, unless it's to the completely software-based 3420 // fp128. 3421 if (Op.getValueType() != MVT::f128) 3422 return Op; 3423 return SDValue(); 3424 } 3425 3426 SDValue AArch64TargetLowering::LowerFSINCOS(SDValue Op, 3427 SelectionDAG &DAG) const { 3428 // For iOS, we want to call an alternative entry point: __sincos_stret, 3429 // which returns the values in two S / D registers. 3430 SDLoc dl(Op); 3431 SDValue Arg = Op.getOperand(0); 3432 EVT ArgVT = Arg.getValueType(); 3433 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext()); 3434 3435 ArgListTy Args; 3436 ArgListEntry Entry; 3437 3438 Entry.Node = Arg; 3439 Entry.Ty = ArgTy; 3440 Entry.IsSExt = false; 3441 Entry.IsZExt = false; 3442 Args.push_back(Entry); 3443 3444 RTLIB::Libcall LC = ArgVT == MVT::f64 ? RTLIB::SINCOS_STRET_F64 3445 : RTLIB::SINCOS_STRET_F32; 3446 const char *LibcallName = getLibcallName(LC); 3447 SDValue Callee = 3448 DAG.getExternalSymbol(LibcallName, getPointerTy(DAG.getDataLayout())); 3449 3450 StructType *RetTy = StructType::get(ArgTy, ArgTy); 3451 TargetLowering::CallLoweringInfo CLI(DAG); 3452 CLI.setDebugLoc(dl) 3453 .setChain(DAG.getEntryNode()) 3454 .setLibCallee(CallingConv::Fast, RetTy, Callee, std::move(Args)); 3455 3456 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI); 3457 return CallResult.first; 3458 } 3459 3460 SDValue AArch64TargetLowering::LowerBITCAST(SDValue Op, 3461 SelectionDAG &DAG) const { 3462 EVT OpVT = Op.getValueType(); 3463 3464 if (useSVEForFixedLengthVectorVT(OpVT)) 3465 return LowerFixedLengthBitcastToSVE(Op, DAG); 3466 3467 if (OpVT != MVT::f16 && OpVT != MVT::bf16) 3468 return SDValue(); 3469 3470 assert(Op.getOperand(0).getValueType() == MVT::i16); 3471 SDLoc DL(Op); 3472 3473 Op = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op.getOperand(0)); 3474 Op = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Op); 3475 return SDValue( 3476 DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL, OpVT, Op, 3477 DAG.getTargetConstant(AArch64::hsub, DL, MVT::i32)), 3478 0); 3479 } 3480 3481 static EVT getExtensionTo64Bits(const EVT &OrigVT) { 3482 if (OrigVT.getSizeInBits() >= 64) 3483 return OrigVT; 3484 3485 assert(OrigVT.isSimple() && "Expecting a simple value type"); 3486 3487 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy; 3488 switch (OrigSimpleTy) { 3489 default: llvm_unreachable("Unexpected Vector Type"); 3490 case MVT::v2i8: 3491 case MVT::v2i16: 3492 return MVT::v2i32; 3493 case MVT::v4i8: 3494 return MVT::v4i16; 3495 } 3496 } 3497 3498 static SDValue addRequiredExtensionForVectorMULL(SDValue N, SelectionDAG &DAG, 3499 const EVT &OrigTy, 3500 const EVT &ExtTy, 3501 unsigned ExtOpcode) { 3502 // The vector originally had a size of OrigTy. It was then extended to ExtTy. 3503 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than 3504 // 64-bits we need to insert a new extension so that it will be 64-bits. 3505 assert(ExtTy.is128BitVector() && "Unexpected extension size"); 3506 if (OrigTy.getSizeInBits() >= 64) 3507 return N; 3508 3509 // Must extend size to at least 64 bits to be used as an operand for VMULL. 3510 EVT NewVT = getExtensionTo64Bits(OrigTy); 3511 3512 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N); 3513 } 3514 3515 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG, 3516 bool isSigned) { 3517 EVT VT = N->getValueType(0); 3518 3519 if (N->getOpcode() != ISD::BUILD_VECTOR) 3520 return false; 3521 3522 for (const SDValue &Elt : N->op_values()) { 3523 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) { 3524 unsigned EltSize = VT.getScalarSizeInBits(); 3525 unsigned HalfSize = EltSize / 2; 3526 if (isSigned) { 3527 if (!isIntN(HalfSize, C->getSExtValue())) 3528 return false; 3529 } else { 3530 if (!isUIntN(HalfSize, C->getZExtValue())) 3531 return false; 3532 } 3533 continue; 3534 } 3535 return false; 3536 } 3537 3538 return true; 3539 } 3540 3541 static SDValue skipExtensionForVectorMULL(SDNode *N, SelectionDAG &DAG) { 3542 if (N->getOpcode() == ISD::SIGN_EXTEND || 3543 N->getOpcode() == ISD::ZERO_EXTEND || N->getOpcode() == ISD::ANY_EXTEND) 3544 return addRequiredExtensionForVectorMULL(N->getOperand(0), DAG, 3545 N->getOperand(0)->getValueType(0), 3546 N->getValueType(0), 3547 N->getOpcode()); 3548 3549 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR"); 3550 EVT VT = N->getValueType(0); 3551 SDLoc dl(N); 3552 unsigned EltSize = VT.getScalarSizeInBits() / 2; 3553 unsigned NumElts = VT.getVectorNumElements(); 3554 MVT TruncVT = MVT::getIntegerVT(EltSize); 3555 SmallVector<SDValue, 8> Ops; 3556 for (unsigned i = 0; i != NumElts; ++i) { 3557 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i)); 3558 const APInt &CInt = C->getAPIntValue(); 3559 // Element types smaller than 32 bits are not legal, so use i32 elements. 3560 // The values are implicitly truncated so sext vs. zext doesn't matter. 3561 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), dl, MVT::i32)); 3562 } 3563 return DAG.getBuildVector(MVT::getVectorVT(TruncVT, NumElts), dl, Ops); 3564 } 3565 3566 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) { 3567 return N->getOpcode() == ISD::SIGN_EXTEND || 3568 N->getOpcode() == ISD::ANY_EXTEND || 3569 isExtendedBUILD_VECTOR(N, DAG, true); 3570 } 3571 3572 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) { 3573 return N->getOpcode() == ISD::ZERO_EXTEND || 3574 N->getOpcode() == ISD::ANY_EXTEND || 3575 isExtendedBUILD_VECTOR(N, DAG, false); 3576 } 3577 3578 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) { 3579 unsigned Opcode = N->getOpcode(); 3580 if (Opcode == ISD::ADD || Opcode == ISD::SUB) { 3581 SDNode *N0 = N->getOperand(0).getNode(); 3582 SDNode *N1 = N->getOperand(1).getNode(); 3583 return N0->hasOneUse() && N1->hasOneUse() && 3584 isSignExtended(N0, DAG) && isSignExtended(N1, DAG); 3585 } 3586 return false; 3587 } 3588 3589 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) { 3590 unsigned Opcode = N->getOpcode(); 3591 if (Opcode == ISD::ADD || Opcode == ISD::SUB) { 3592 SDNode *N0 = N->getOperand(0).getNode(); 3593 SDNode *N1 = N->getOperand(1).getNode(); 3594 return N0->hasOneUse() && N1->hasOneUse() && 3595 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG); 3596 } 3597 return false; 3598 } 3599 3600 SDValue AArch64TargetLowering::LowerFLT_ROUNDS_(SDValue Op, 3601 SelectionDAG &DAG) const { 3602 // The rounding mode is in bits 23:22 of the FPSCR. 3603 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0 3604 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3) 3605 // so that the shift + and get folded into a bitfield extract. 3606 SDLoc dl(Op); 3607 3608 SDValue Chain = Op.getOperand(0); 3609 SDValue FPCR_64 = DAG.getNode( 3610 ISD::INTRINSIC_W_CHAIN, dl, {MVT::i64, MVT::Other}, 3611 {Chain, DAG.getConstant(Intrinsic::aarch64_get_fpcr, dl, MVT::i64)}); 3612 Chain = FPCR_64.getValue(1); 3613 SDValue FPCR_32 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, FPCR_64); 3614 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPCR_32, 3615 DAG.getConstant(1U << 22, dl, MVT::i32)); 3616 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds, 3617 DAG.getConstant(22, dl, MVT::i32)); 3618 SDValue AND = DAG.getNode(ISD::AND, dl, MVT::i32, RMODE, 3619 DAG.getConstant(3, dl, MVT::i32)); 3620 return DAG.getMergeValues({AND, Chain}, dl); 3621 } 3622 3623 SDValue AArch64TargetLowering::LowerSET_ROUNDING(SDValue Op, 3624 SelectionDAG &DAG) const { 3625 SDLoc DL(Op); 3626 SDValue Chain = Op->getOperand(0); 3627 SDValue RMValue = Op->getOperand(1); 3628 3629 // The rounding mode is in bits 23:22 of the FPCR. 3630 // The llvm.set.rounding argument value to the rounding mode in FPCR mapping 3631 // is 0->3, 1->0, 2->1, 3->2. The formula we use to implement this is 3632 // ((arg - 1) & 3) << 22). 3633 // 3634 // The argument of llvm.set.rounding must be within the segment [0, 3], so 3635 // NearestTiesToAway (4) is not handled here. It is responsibility of the code 3636 // generated llvm.set.rounding to ensure this condition. 3637 3638 // Calculate new value of FPCR[23:22]. 3639 RMValue = DAG.getNode(ISD::SUB, DL, MVT::i32, RMValue, 3640 DAG.getConstant(1, DL, MVT::i32)); 3641 RMValue = DAG.getNode(ISD::AND, DL, MVT::i32, RMValue, 3642 DAG.getConstant(0x3, DL, MVT::i32)); 3643 RMValue = 3644 DAG.getNode(ISD::SHL, DL, MVT::i32, RMValue, 3645 DAG.getConstant(AArch64::RoundingBitsPos, DL, MVT::i32)); 3646 RMValue = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, RMValue); 3647 3648 // Get current value of FPCR. 3649 SDValue Ops[] = { 3650 Chain, DAG.getTargetConstant(Intrinsic::aarch64_get_fpcr, DL, MVT::i64)}; 3651 SDValue FPCR = 3652 DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL, {MVT::i64, MVT::Other}, Ops); 3653 Chain = FPCR.getValue(1); 3654 FPCR = FPCR.getValue(0); 3655 3656 // Put new rounding mode into FPSCR[23:22]. 3657 const int RMMask = ~(AArch64::Rounding::rmMask << AArch64::RoundingBitsPos); 3658 FPCR = DAG.getNode(ISD::AND, DL, MVT::i64, FPCR, 3659 DAG.getConstant(RMMask, DL, MVT::i64)); 3660 FPCR = DAG.getNode(ISD::OR, DL, MVT::i64, FPCR, RMValue); 3661 SDValue Ops2[] = { 3662 Chain, DAG.getTargetConstant(Intrinsic::aarch64_set_fpcr, DL, MVT::i64), 3663 FPCR}; 3664 return DAG.getNode(ISD::INTRINSIC_VOID, DL, MVT::Other, Ops2); 3665 } 3666 3667 SDValue AArch64TargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const { 3668 EVT VT = Op.getValueType(); 3669 3670 // If SVE is available then i64 vector multiplications can also be made legal. 3671 bool OverrideNEON = VT == MVT::v2i64 || VT == MVT::v1i64; 3672 3673 if (VT.isScalableVector() || useSVEForFixedLengthVectorVT(VT, OverrideNEON)) 3674 return LowerToPredicatedOp(Op, DAG, AArch64ISD::MUL_PRED, OverrideNEON); 3675 3676 // Multiplications are only custom-lowered for 128-bit vectors so that 3677 // VMULL can be detected. Otherwise v2i64 multiplications are not legal. 3678 assert(VT.is128BitVector() && VT.isInteger() && 3679 "unexpected type for custom-lowering ISD::MUL"); 3680 SDNode *N0 = Op.getOperand(0).getNode(); 3681 SDNode *N1 = Op.getOperand(1).getNode(); 3682 unsigned NewOpc = 0; 3683 bool isMLA = false; 3684 bool isN0SExt = isSignExtended(N0, DAG); 3685 bool isN1SExt = isSignExtended(N1, DAG); 3686 if (isN0SExt && isN1SExt) 3687 NewOpc = AArch64ISD::SMULL; 3688 else { 3689 bool isN0ZExt = isZeroExtended(N0, DAG); 3690 bool isN1ZExt = isZeroExtended(N1, DAG); 3691 if (isN0ZExt && isN1ZExt) 3692 NewOpc = AArch64ISD::UMULL; 3693 else if (isN1SExt || isN1ZExt) { 3694 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these 3695 // into (s/zext A * s/zext C) + (s/zext B * s/zext C) 3696 if (isN1SExt && isAddSubSExt(N0, DAG)) { 3697 NewOpc = AArch64ISD::SMULL; 3698 isMLA = true; 3699 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) { 3700 NewOpc = AArch64ISD::UMULL; 3701 isMLA = true; 3702 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) { 3703 std::swap(N0, N1); 3704 NewOpc = AArch64ISD::UMULL; 3705 isMLA = true; 3706 } 3707 } 3708 3709 if (!NewOpc) { 3710 if (VT == MVT::v2i64) 3711 // Fall through to expand this. It is not legal. 3712 return SDValue(); 3713 else 3714 // Other vector multiplications are legal. 3715 return Op; 3716 } 3717 } 3718 3719 // Legalize to a S/UMULL instruction 3720 SDLoc DL(Op); 3721 SDValue Op0; 3722 SDValue Op1 = skipExtensionForVectorMULL(N1, DAG); 3723 if (!isMLA) { 3724 Op0 = skipExtensionForVectorMULL(N0, DAG); 3725 assert(Op0.getValueType().is64BitVector() && 3726 Op1.getValueType().is64BitVector() && 3727 "unexpected types for extended operands to VMULL"); 3728 return DAG.getNode(NewOpc, DL, VT, Op0, Op1); 3729 } 3730 // Optimizing (zext A + zext B) * C, to (S/UMULL A, C) + (S/UMULL B, C) during 3731 // isel lowering to take advantage of no-stall back to back s/umul + s/umla. 3732 // This is true for CPUs with accumulate forwarding such as Cortex-A53/A57 3733 SDValue N00 = skipExtensionForVectorMULL(N0->getOperand(0).getNode(), DAG); 3734 SDValue N01 = skipExtensionForVectorMULL(N0->getOperand(1).getNode(), DAG); 3735 EVT Op1VT = Op1.getValueType(); 3736 return DAG.getNode(N0->getOpcode(), DL, VT, 3737 DAG.getNode(NewOpc, DL, VT, 3738 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1), 3739 DAG.getNode(NewOpc, DL, VT, 3740 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1)); 3741 } 3742 3743 static inline SDValue getPTrue(SelectionDAG &DAG, SDLoc DL, EVT VT, 3744 int Pattern) { 3745 return DAG.getNode(AArch64ISD::PTRUE, DL, VT, 3746 DAG.getTargetConstant(Pattern, DL, MVT::i32)); 3747 } 3748 3749 static SDValue lowerConvertToSVBool(SDValue Op, SelectionDAG &DAG) { 3750 SDLoc DL(Op); 3751 EVT OutVT = Op.getValueType(); 3752 SDValue InOp = Op.getOperand(1); 3753 EVT InVT = InOp.getValueType(); 3754 3755 // Return the operand if the cast isn't changing type, 3756 // i.e. <n x 16 x i1> -> <n x 16 x i1> 3757 if (InVT == OutVT) 3758 return InOp; 3759 3760 SDValue Reinterpret = 3761 DAG.getNode(AArch64ISD::REINTERPRET_CAST, DL, OutVT, InOp); 3762 3763 // If the argument converted to an svbool is a ptrue or a comparison, the 3764 // lanes introduced by the widening are zero by construction. 3765 switch (InOp.getOpcode()) { 3766 case AArch64ISD::SETCC_MERGE_ZERO: 3767 return Reinterpret; 3768 case ISD::INTRINSIC_WO_CHAIN: 3769 if (InOp.getConstantOperandVal(0) == Intrinsic::aarch64_sve_ptrue) 3770 return Reinterpret; 3771 } 3772 3773 // Otherwise, zero the newly introduced lanes. 3774 SDValue Mask = getPTrue(DAG, DL, InVT, AArch64SVEPredPattern::all); 3775 SDValue MaskReinterpret = 3776 DAG.getNode(AArch64ISD::REINTERPRET_CAST, DL, OutVT, Mask); 3777 return DAG.getNode(ISD::AND, DL, OutVT, Reinterpret, MaskReinterpret); 3778 } 3779 3780 SDValue AArch64TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, 3781 SelectionDAG &DAG) const { 3782 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 3783 SDLoc dl(Op); 3784 switch (IntNo) { 3785 default: return SDValue(); // Don't custom lower most intrinsics. 3786 case Intrinsic::thread_pointer: { 3787 EVT PtrVT = getPointerTy(DAG.getDataLayout()); 3788 return DAG.getNode(AArch64ISD::THREAD_POINTER, dl, PtrVT); 3789 } 3790 case Intrinsic::aarch64_neon_abs: { 3791 EVT Ty = Op.getValueType(); 3792 if (Ty == MVT::i64) { 3793 SDValue Result = DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, 3794 Op.getOperand(1)); 3795 Result = DAG.getNode(ISD::ABS, dl, MVT::v1i64, Result); 3796 return DAG.getNode(ISD::BITCAST, dl, MVT::i64, Result); 3797 } else if (Ty.isVector() && Ty.isInteger() && isTypeLegal(Ty)) { 3798 return DAG.getNode(ISD::ABS, dl, Ty, Op.getOperand(1)); 3799 } else { 3800 report_fatal_error("Unexpected type for AArch64 NEON intrinic"); 3801 } 3802 } 3803 case Intrinsic::aarch64_neon_smax: 3804 return DAG.getNode(ISD::SMAX, dl, Op.getValueType(), 3805 Op.getOperand(1), Op.getOperand(2)); 3806 case Intrinsic::aarch64_neon_umax: 3807 return DAG.getNode(ISD::UMAX, dl, Op.getValueType(), 3808 Op.getOperand(1), Op.getOperand(2)); 3809 case Intrinsic::aarch64_neon_smin: 3810 return DAG.getNode(ISD::SMIN, dl, Op.getValueType(), 3811 Op.getOperand(1), Op.getOperand(2)); 3812 case Intrinsic::aarch64_neon_umin: 3813 return DAG.getNode(ISD::UMIN, dl, Op.getValueType(), 3814 Op.getOperand(1), Op.getOperand(2)); 3815 3816 case Intrinsic::aarch64_sve_sunpkhi: 3817 return DAG.getNode(AArch64ISD::SUNPKHI, dl, Op.getValueType(), 3818 Op.getOperand(1)); 3819 case Intrinsic::aarch64_sve_sunpklo: 3820 return DAG.getNode(AArch64ISD::SUNPKLO, dl, Op.getValueType(), 3821 Op.getOperand(1)); 3822 case Intrinsic::aarch64_sve_uunpkhi: 3823 return DAG.getNode(AArch64ISD::UUNPKHI, dl, Op.getValueType(), 3824 Op.getOperand(1)); 3825 case Intrinsic::aarch64_sve_uunpklo: 3826 return DAG.getNode(AArch64ISD::UUNPKLO, dl, Op.getValueType(), 3827 Op.getOperand(1)); 3828 case Intrinsic::aarch64_sve_clasta_n: 3829 return DAG.getNode(AArch64ISD::CLASTA_N, dl, Op.getValueType(), 3830 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3)); 3831 case Intrinsic::aarch64_sve_clastb_n: 3832 return DAG.getNode(AArch64ISD::CLASTB_N, dl, Op.getValueType(), 3833 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3)); 3834 case Intrinsic::aarch64_sve_lasta: 3835 return DAG.getNode(AArch64ISD::LASTA, dl, Op.getValueType(), 3836 Op.getOperand(1), Op.getOperand(2)); 3837 case Intrinsic::aarch64_sve_lastb: 3838 return DAG.getNode(AArch64ISD::LASTB, dl, Op.getValueType(), 3839 Op.getOperand(1), Op.getOperand(2)); 3840 case Intrinsic::aarch64_sve_rev: 3841 return DAG.getNode(ISD::VECTOR_REVERSE, dl, Op.getValueType(), 3842 Op.getOperand(1)); 3843 case Intrinsic::aarch64_sve_tbl: 3844 return DAG.getNode(AArch64ISD::TBL, dl, Op.getValueType(), 3845 Op.getOperand(1), Op.getOperand(2)); 3846 case Intrinsic::aarch64_sve_trn1: 3847 return DAG.getNode(AArch64ISD::TRN1, dl, Op.getValueType(), 3848 Op.getOperand(1), Op.getOperand(2)); 3849 case Intrinsic::aarch64_sve_trn2: 3850 return DAG.getNode(AArch64ISD::TRN2, dl, Op.getValueType(), 3851 Op.getOperand(1), Op.getOperand(2)); 3852 case Intrinsic::aarch64_sve_uzp1: 3853 return DAG.getNode(AArch64ISD::UZP1, dl, Op.getValueType(), 3854 Op.getOperand(1), Op.getOperand(2)); 3855 case Intrinsic::aarch64_sve_uzp2: 3856 return DAG.getNode(AArch64ISD::UZP2, dl, Op.getValueType(), 3857 Op.getOperand(1), Op.getOperand(2)); 3858 case Intrinsic::aarch64_sve_zip1: 3859 return DAG.getNode(AArch64ISD::ZIP1, dl, Op.getValueType(), 3860 Op.getOperand(1), Op.getOperand(2)); 3861 case Intrinsic::aarch64_sve_zip2: 3862 return DAG.getNode(AArch64ISD::ZIP2, dl, Op.getValueType(), 3863 Op.getOperand(1), Op.getOperand(2)); 3864 case Intrinsic::aarch64_sve_ptrue: 3865 return DAG.getNode(AArch64ISD::PTRUE, dl, Op.getValueType(), 3866 Op.getOperand(1)); 3867 case Intrinsic::aarch64_sve_clz: 3868 return DAG.getNode(AArch64ISD::CTLZ_MERGE_PASSTHRU, dl, Op.getValueType(), 3869 Op.getOperand(2), Op.getOperand(3), Op.getOperand(1)); 3870 case Intrinsic::aarch64_sve_cnt: { 3871 SDValue Data = Op.getOperand(3); 3872 // CTPOP only supports integer operands. 3873 if (Data.getValueType().isFloatingPoint()) 3874 Data = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Data); 3875 return DAG.getNode(AArch64ISD::CTPOP_MERGE_PASSTHRU, dl, Op.getValueType(), 3876 Op.getOperand(2), Data, Op.getOperand(1)); 3877 } 3878 case Intrinsic::aarch64_sve_dupq_lane: 3879 return LowerDUPQLane(Op, DAG); 3880 case Intrinsic::aarch64_sve_convert_from_svbool: 3881 return DAG.getNode(AArch64ISD::REINTERPRET_CAST, dl, Op.getValueType(), 3882 Op.getOperand(1)); 3883 case Intrinsic::aarch64_sve_convert_to_svbool: 3884 return lowerConvertToSVBool(Op, DAG); 3885 case Intrinsic::aarch64_sve_fneg: 3886 return DAG.getNode(AArch64ISD::FNEG_MERGE_PASSTHRU, dl, Op.getValueType(), 3887 Op.getOperand(2), Op.getOperand(3), Op.getOperand(1)); 3888 case Intrinsic::aarch64_sve_frintp: 3889 return DAG.getNode(AArch64ISD::FCEIL_MERGE_PASSTHRU, dl, Op.getValueType(), 3890 Op.getOperand(2), Op.getOperand(3), Op.getOperand(1)); 3891 case Intrinsic::aarch64_sve_frintm: 3892 return DAG.getNode(AArch64ISD::FFLOOR_MERGE_PASSTHRU, dl, Op.getValueType(), 3893 Op.getOperand(2), Op.getOperand(3), Op.getOperand(1)); 3894 case Intrinsic::aarch64_sve_frinti: 3895 return DAG.getNode(AArch64ISD::FNEARBYINT_MERGE_PASSTHRU, dl, Op.getValueType(), 3896 Op.getOperand(2), Op.getOperand(3), Op.getOperand(1)); 3897 case Intrinsic::aarch64_sve_frintx: 3898 return DAG.getNode(AArch64ISD::FRINT_MERGE_PASSTHRU, dl, Op.getValueType(), 3899 Op.getOperand(2), Op.getOperand(3), Op.getOperand(1)); 3900 case Intrinsic::aarch64_sve_frinta: 3901 return DAG.getNode(AArch64ISD::FROUND_MERGE_PASSTHRU, dl, Op.getValueType(), 3902 Op.getOperand(2), Op.getOperand(3), Op.getOperand(1)); 3903 case Intrinsic::aarch64_sve_frintn: 3904 return DAG.getNode(AArch64ISD::FROUNDEVEN_MERGE_PASSTHRU, dl, Op.getValueType(), 3905 Op.getOperand(2), Op.getOperand(3), Op.getOperand(1)); 3906 case Intrinsic::aarch64_sve_frintz: 3907 return DAG.getNode(AArch64ISD::FTRUNC_MERGE_PASSTHRU, dl, Op.getValueType(), 3908 Op.getOperand(2), Op.getOperand(3), Op.getOperand(1)); 3909 case Intrinsic::aarch64_sve_ucvtf: 3910 return DAG.getNode(AArch64ISD::UINT_TO_FP_MERGE_PASSTHRU, dl, 3911 Op.getValueType(), Op.getOperand(2), Op.getOperand(3), 3912 Op.getOperand(1)); 3913 case Intrinsic::aarch64_sve_scvtf: 3914 return DAG.getNode(AArch64ISD::SINT_TO_FP_MERGE_PASSTHRU, dl, 3915 Op.getValueType(), Op.getOperand(2), Op.getOperand(3), 3916 Op.getOperand(1)); 3917 case Intrinsic::aarch64_sve_fcvtzu: 3918 return DAG.getNode(AArch64ISD::FCVTZU_MERGE_PASSTHRU, dl, 3919 Op.getValueType(), Op.getOperand(2), Op.getOperand(3), 3920 Op.getOperand(1)); 3921 case Intrinsic::aarch64_sve_fcvtzs: 3922 return DAG.getNode(AArch64ISD::FCVTZS_MERGE_PASSTHRU, dl, 3923 Op.getValueType(), Op.getOperand(2), Op.getOperand(3), 3924 Op.getOperand(1)); 3925 case Intrinsic::aarch64_sve_fsqrt: 3926 return DAG.getNode(AArch64ISD::FSQRT_MERGE_PASSTHRU, dl, Op.getValueType(), 3927 Op.getOperand(2), Op.getOperand(3), Op.getOperand(1)); 3928 case Intrinsic::aarch64_sve_frecpx: 3929 return DAG.getNode(AArch64ISD::FRECPX_MERGE_PASSTHRU, dl, Op.getValueType(), 3930 Op.getOperand(2), Op.getOperand(3), Op.getOperand(1)); 3931 case Intrinsic::aarch64_sve_fabs: 3932 return DAG.getNode(AArch64ISD::FABS_MERGE_PASSTHRU, dl, Op.getValueType(), 3933 Op.getOperand(2), Op.getOperand(3), Op.getOperand(1)); 3934 case Intrinsic::aarch64_sve_abs: 3935 return DAG.getNode(AArch64ISD::ABS_MERGE_PASSTHRU, dl, Op.getValueType(), 3936 Op.getOperand(2), Op.getOperand(3), Op.getOperand(1)); 3937 case Intrinsic::aarch64_sve_neg: 3938 return DAG.getNode(AArch64ISD::NEG_MERGE_PASSTHRU, dl, Op.getValueType(), 3939 Op.getOperand(2), Op.getOperand(3), Op.getOperand(1)); 3940 case Intrinsic::aarch64_sve_insr: { 3941 SDValue Scalar = Op.getOperand(2); 3942 EVT ScalarTy = Scalar.getValueType(); 3943 if ((ScalarTy == MVT::i8) || (ScalarTy == MVT::i16)) 3944 Scalar = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Scalar); 3945 3946 return DAG.getNode(AArch64ISD::INSR, dl, Op.getValueType(), 3947 Op.getOperand(1), Scalar); 3948 } 3949 case Intrinsic::aarch64_sve_rbit: 3950 return DAG.getNode(AArch64ISD::BITREVERSE_MERGE_PASSTHRU, dl, 3951 Op.getValueType(), Op.getOperand(2), Op.getOperand(3), 3952 Op.getOperand(1)); 3953 case Intrinsic::aarch64_sve_revb: 3954 return DAG.getNode(AArch64ISD::BSWAP_MERGE_PASSTHRU, dl, Op.getValueType(), 3955 Op.getOperand(2), Op.getOperand(3), Op.getOperand(1)); 3956 case Intrinsic::aarch64_sve_sxtb: 3957 return DAG.getNode( 3958 AArch64ISD::SIGN_EXTEND_INREG_MERGE_PASSTHRU, dl, Op.getValueType(), 3959 Op.getOperand(2), Op.getOperand(3), 3960 DAG.getValueType(Op.getValueType().changeVectorElementType(MVT::i8)), 3961 Op.getOperand(1)); 3962 case Intrinsic::aarch64_sve_sxth: 3963 return DAG.getNode( 3964 AArch64ISD::SIGN_EXTEND_INREG_MERGE_PASSTHRU, dl, Op.getValueType(), 3965 Op.getOperand(2), Op.getOperand(3), 3966 DAG.getValueType(Op.getValueType().changeVectorElementType(MVT::i16)), 3967 Op.getOperand(1)); 3968 case Intrinsic::aarch64_sve_sxtw: 3969 return DAG.getNode( 3970 AArch64ISD::SIGN_EXTEND_INREG_MERGE_PASSTHRU, dl, Op.getValueType(), 3971 Op.getOperand(2), Op.getOperand(3), 3972 DAG.getValueType(Op.getValueType().changeVectorElementType(MVT::i32)), 3973 Op.getOperand(1)); 3974 case Intrinsic::aarch64_sve_uxtb: 3975 return DAG.getNode( 3976 AArch64ISD::ZERO_EXTEND_INREG_MERGE_PASSTHRU, dl, Op.getValueType(), 3977 Op.getOperand(2), Op.getOperand(3), 3978 DAG.getValueType(Op.getValueType().changeVectorElementType(MVT::i8)), 3979 Op.getOperand(1)); 3980 case Intrinsic::aarch64_sve_uxth: 3981 return DAG.getNode( 3982 AArch64ISD::ZERO_EXTEND_INREG_MERGE_PASSTHRU, dl, Op.getValueType(), 3983 Op.getOperand(2), Op.getOperand(3), 3984 DAG.getValueType(Op.getValueType().changeVectorElementType(MVT::i16)), 3985 Op.getOperand(1)); 3986 case Intrinsic::aarch64_sve_uxtw: 3987 return DAG.getNode( 3988 AArch64ISD::ZERO_EXTEND_INREG_MERGE_PASSTHRU, dl, Op.getValueType(), 3989 Op.getOperand(2), Op.getOperand(3), 3990 DAG.getValueType(Op.getValueType().changeVectorElementType(MVT::i32)), 3991 Op.getOperand(1)); 3992 3993 case Intrinsic::localaddress: { 3994 const auto &MF = DAG.getMachineFunction(); 3995 const auto *RegInfo = Subtarget->getRegisterInfo(); 3996 unsigned Reg = RegInfo->getLocalAddressRegister(MF); 3997 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, 3998 Op.getSimpleValueType()); 3999 } 4000 4001 case Intrinsic::eh_recoverfp: { 4002 // FIXME: This needs to be implemented to correctly handle highly aligned 4003 // stack objects. For now we simply return the incoming FP. Refer D53541 4004 // for more details. 4005 SDValue FnOp = Op.getOperand(1); 4006 SDValue IncomingFPOp = Op.getOperand(2); 4007 GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp); 4008 auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr); 4009 if (!Fn) 4010 report_fatal_error( 4011 "llvm.eh.recoverfp must take a function as the first argument"); 4012 return IncomingFPOp; 4013 } 4014 4015 case Intrinsic::aarch64_neon_vsri: 4016 case Intrinsic::aarch64_neon_vsli: { 4017 EVT Ty = Op.getValueType(); 4018 4019 if (!Ty.isVector()) 4020 report_fatal_error("Unexpected type for aarch64_neon_vsli"); 4021 4022 assert(Op.getConstantOperandVal(3) <= Ty.getScalarSizeInBits()); 4023 4024 bool IsShiftRight = IntNo == Intrinsic::aarch64_neon_vsri; 4025 unsigned Opcode = IsShiftRight ? AArch64ISD::VSRI : AArch64ISD::VSLI; 4026 return DAG.getNode(Opcode, dl, Ty, Op.getOperand(1), Op.getOperand(2), 4027 Op.getOperand(3)); 4028 } 4029 4030 case Intrinsic::aarch64_neon_srhadd: 4031 case Intrinsic::aarch64_neon_urhadd: 4032 case Intrinsic::aarch64_neon_shadd: 4033 case Intrinsic::aarch64_neon_uhadd: { 4034 bool IsSignedAdd = (IntNo == Intrinsic::aarch64_neon_srhadd || 4035 IntNo == Intrinsic::aarch64_neon_shadd); 4036 bool IsRoundingAdd = (IntNo == Intrinsic::aarch64_neon_srhadd || 4037 IntNo == Intrinsic::aarch64_neon_urhadd); 4038 unsigned Opcode = 4039 IsSignedAdd ? (IsRoundingAdd ? AArch64ISD::SRHADD : AArch64ISD::SHADD) 4040 : (IsRoundingAdd ? AArch64ISD::URHADD : AArch64ISD::UHADD); 4041 return DAG.getNode(Opcode, dl, Op.getValueType(), Op.getOperand(1), 4042 Op.getOperand(2)); 4043 } 4044 case Intrinsic::aarch64_neon_sabd: 4045 case Intrinsic::aarch64_neon_uabd: { 4046 unsigned Opcode = IntNo == Intrinsic::aarch64_neon_uabd ? AArch64ISD::UABD 4047 : AArch64ISD::SABD; 4048 return DAG.getNode(Opcode, dl, Op.getValueType(), Op.getOperand(1), 4049 Op.getOperand(2)); 4050 } 4051 case Intrinsic::aarch64_neon_sdot: 4052 case Intrinsic::aarch64_neon_udot: 4053 case Intrinsic::aarch64_sve_sdot: 4054 case Intrinsic::aarch64_sve_udot: { 4055 unsigned Opcode = (IntNo == Intrinsic::aarch64_neon_udot || 4056 IntNo == Intrinsic::aarch64_sve_udot) 4057 ? AArch64ISD::UDOT 4058 : AArch64ISD::SDOT; 4059 return DAG.getNode(Opcode, dl, Op.getValueType(), Op.getOperand(1), 4060 Op.getOperand(2), Op.getOperand(3)); 4061 } 4062 } 4063 } 4064 4065 bool AArch64TargetLowering::shouldExtendGSIndex(EVT VT, EVT &EltTy) const { 4066 if (VT.getVectorElementType() == MVT::i8 || 4067 VT.getVectorElementType() == MVT::i16) { 4068 EltTy = MVT::i32; 4069 return true; 4070 } 4071 return false; 4072 } 4073 4074 bool AArch64TargetLowering::shouldRemoveExtendFromGSIndex(EVT VT) const { 4075 if (VT.getVectorElementType() == MVT::i32 && 4076 VT.getVectorElementCount().getKnownMinValue() >= 4) 4077 return true; 4078 4079 return false; 4080 } 4081 4082 bool AArch64TargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const { 4083 return ExtVal.getValueType().isScalableVector(); 4084 } 4085 4086 unsigned getGatherVecOpcode(bool IsScaled, bool IsSigned, bool NeedsExtend) { 4087 std::map<std::tuple<bool, bool, bool>, unsigned> AddrModes = { 4088 {std::make_tuple(/*Scaled*/ false, /*Signed*/ false, /*Extend*/ false), 4089 AArch64ISD::GLD1_MERGE_ZERO}, 4090 {std::make_tuple(/*Scaled*/ false, /*Signed*/ false, /*Extend*/ true), 4091 AArch64ISD::GLD1_UXTW_MERGE_ZERO}, 4092 {std::make_tuple(/*Scaled*/ false, /*Signed*/ true, /*Extend*/ false), 4093 AArch64ISD::GLD1_MERGE_ZERO}, 4094 {std::make_tuple(/*Scaled*/ false, /*Signed*/ true, /*Extend*/ true), 4095 AArch64ISD::GLD1_SXTW_MERGE_ZERO}, 4096 {std::make_tuple(/*Scaled*/ true, /*Signed*/ false, /*Extend*/ false), 4097 AArch64ISD::GLD1_SCALED_MERGE_ZERO}, 4098 {std::make_tuple(/*Scaled*/ true, /*Signed*/ false, /*Extend*/ true), 4099 AArch64ISD::GLD1_UXTW_SCALED_MERGE_ZERO}, 4100 {std::make_tuple(/*Scaled*/ true, /*Signed*/ true, /*Extend*/ false), 4101 AArch64ISD::GLD1_SCALED_MERGE_ZERO}, 4102 {std::make_tuple(/*Scaled*/ true, /*Signed*/ true, /*Extend*/ true), 4103 AArch64ISD::GLD1_SXTW_SCALED_MERGE_ZERO}, 4104 }; 4105 auto Key = std::make_tuple(IsScaled, IsSigned, NeedsExtend); 4106 return AddrModes.find(Key)->second; 4107 } 4108 4109 unsigned getScatterVecOpcode(bool IsScaled, bool IsSigned, bool NeedsExtend) { 4110 std::map<std::tuple<bool, bool, bool>, unsigned> AddrModes = { 4111 {std::make_tuple(/*Scaled*/ false, /*Signed*/ false, /*Extend*/ false), 4112 AArch64ISD::SST1_PRED}, 4113 {std::make_tuple(/*Scaled*/ false, /*Signed*/ false, /*Extend*/ true), 4114 AArch64ISD::SST1_UXTW_PRED}, 4115 {std::make_tuple(/*Scaled*/ false, /*Signed*/ true, /*Extend*/ false), 4116 AArch64ISD::SST1_PRED}, 4117 {std::make_tuple(/*Scaled*/ false, /*Signed*/ true, /*Extend*/ true), 4118 AArch64ISD::SST1_SXTW_PRED}, 4119 {std::make_tuple(/*Scaled*/ true, /*Signed*/ false, /*Extend*/ false), 4120 AArch64ISD::SST1_SCALED_PRED}, 4121 {std::make_tuple(/*Scaled*/ true, /*Signed*/ false, /*Extend*/ true), 4122 AArch64ISD::SST1_UXTW_SCALED_PRED}, 4123 {std::make_tuple(/*Scaled*/ true, /*Signed*/ true, /*Extend*/ false), 4124 AArch64ISD::SST1_SCALED_PRED}, 4125 {std::make_tuple(/*Scaled*/ true, /*Signed*/ true, /*Extend*/ true), 4126 AArch64ISD::SST1_SXTW_SCALED_PRED}, 4127 }; 4128 auto Key = std::make_tuple(IsScaled, IsSigned, NeedsExtend); 4129 return AddrModes.find(Key)->second; 4130 } 4131 4132 unsigned getSignExtendedGatherOpcode(unsigned Opcode) { 4133 switch (Opcode) { 4134 default: 4135 llvm_unreachable("unimplemented opcode"); 4136 return Opcode; 4137 case AArch64ISD::GLD1_MERGE_ZERO: 4138 return AArch64ISD::GLD1S_MERGE_ZERO; 4139 case AArch64ISD::GLD1_IMM_MERGE_ZERO: 4140 return AArch64ISD::GLD1S_IMM_MERGE_ZERO; 4141 case AArch64ISD::GLD1_UXTW_MERGE_ZERO: 4142 return AArch64ISD::GLD1S_UXTW_MERGE_ZERO; 4143 case AArch64ISD::GLD1_SXTW_MERGE_ZERO: 4144 return AArch64ISD::GLD1S_SXTW_MERGE_ZERO; 4145 case AArch64ISD::GLD1_SCALED_MERGE_ZERO: 4146 return AArch64ISD::GLD1S_SCALED_MERGE_ZERO; 4147 case AArch64ISD::GLD1_UXTW_SCALED_MERGE_ZERO: 4148 return AArch64ISD::GLD1S_UXTW_SCALED_MERGE_ZERO; 4149 case AArch64ISD::GLD1_SXTW_SCALED_MERGE_ZERO: 4150 return AArch64ISD::GLD1S_SXTW_SCALED_MERGE_ZERO; 4151 } 4152 } 4153 4154 bool getGatherScatterIndexIsExtended(SDValue Index) { 4155 unsigned Opcode = Index.getOpcode(); 4156 if (Opcode == ISD::SIGN_EXTEND_INREG) 4157 return true; 4158 4159 if (Opcode == ISD::AND) { 4160 SDValue Splat = Index.getOperand(1); 4161 if (Splat.getOpcode() != ISD::SPLAT_VECTOR) 4162 return false; 4163 ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(Splat.getOperand(0)); 4164 if (!Mask || Mask->getZExtValue() != 0xFFFFFFFF) 4165 return false; 4166 return true; 4167 } 4168 4169 return false; 4170 } 4171 4172 // If the base pointer of a masked gather or scatter is null, we 4173 // may be able to swap BasePtr & Index and use the vector + register 4174 // or vector + immediate addressing mode, e.g. 4175 // VECTOR + REGISTER: 4176 // getelementptr nullptr, <vscale x N x T> (splat(%offset)) + %indices) 4177 // -> getelementptr %offset, <vscale x N x T> %indices 4178 // VECTOR + IMMEDIATE: 4179 // getelementptr nullptr, <vscale x N x T> (splat(#x)) + %indices) 4180 // -> getelementptr #x, <vscale x N x T> %indices 4181 void selectGatherScatterAddrMode(SDValue &BasePtr, SDValue &Index, EVT MemVT, 4182 unsigned &Opcode, bool IsGather, 4183 SelectionDAG &DAG) { 4184 if (!isNullConstant(BasePtr)) 4185 return; 4186 4187 ConstantSDNode *Offset = nullptr; 4188 if (Index.getOpcode() == ISD::ADD) 4189 if (auto SplatVal = DAG.getSplatValue(Index.getOperand(1))) { 4190 if (isa<ConstantSDNode>(SplatVal)) 4191 Offset = cast<ConstantSDNode>(SplatVal); 4192 else { 4193 BasePtr = SplatVal; 4194 Index = Index->getOperand(0); 4195 return; 4196 } 4197 } 4198 4199 unsigned NewOp = 4200 IsGather ? AArch64ISD::GLD1_IMM_MERGE_ZERO : AArch64ISD::SST1_IMM_PRED; 4201 4202 if (!Offset) { 4203 std::swap(BasePtr, Index); 4204 Opcode = NewOp; 4205 return; 4206 } 4207 4208 uint64_t OffsetVal = Offset->getZExtValue(); 4209 unsigned ScalarSizeInBytes = MemVT.getScalarSizeInBits() / 8; 4210 auto ConstOffset = DAG.getConstant(OffsetVal, SDLoc(Index), MVT::i64); 4211 4212 if (OffsetVal % ScalarSizeInBytes || OffsetVal / ScalarSizeInBytes > 31) { 4213 // Index is out of range for the immediate addressing mode 4214 BasePtr = ConstOffset; 4215 Index = Index->getOperand(0); 4216 return; 4217 } 4218 4219 // Immediate is in range 4220 Opcode = NewOp; 4221 BasePtr = Index->getOperand(0); 4222 Index = ConstOffset; 4223 } 4224 4225 SDValue AArch64TargetLowering::LowerMGATHER(SDValue Op, 4226 SelectionDAG &DAG) const { 4227 SDLoc DL(Op); 4228 MaskedGatherSDNode *MGT = cast<MaskedGatherSDNode>(Op); 4229 assert(MGT && "Can only custom lower gather load nodes"); 4230 4231 SDValue Index = MGT->getIndex(); 4232 SDValue Chain = MGT->getChain(); 4233 SDValue PassThru = MGT->getPassThru(); 4234 SDValue Mask = MGT->getMask(); 4235 SDValue BasePtr = MGT->getBasePtr(); 4236 ISD::LoadExtType ExtTy = MGT->getExtensionType(); 4237 4238 ISD::MemIndexType IndexType = MGT->getIndexType(); 4239 bool IsScaled = 4240 IndexType == ISD::SIGNED_SCALED || IndexType == ISD::UNSIGNED_SCALED; 4241 bool IsSigned = 4242 IndexType == ISD::SIGNED_SCALED || IndexType == ISD::SIGNED_UNSCALED; 4243 bool IdxNeedsExtend = 4244 getGatherScatterIndexIsExtended(Index) || 4245 Index.getSimpleValueType().getVectorElementType() == MVT::i32; 4246 bool ResNeedsSignExtend = ExtTy == ISD::EXTLOAD || ExtTy == ISD::SEXTLOAD; 4247 4248 EVT VT = PassThru.getSimpleValueType(); 4249 EVT MemVT = MGT->getMemoryVT(); 4250 SDValue InputVT = DAG.getValueType(MemVT); 4251 4252 if (VT.getVectorElementType() == MVT::bf16 && 4253 !static_cast<const AArch64Subtarget &>(DAG.getSubtarget()).hasBF16()) 4254 return SDValue(); 4255 4256 // Handle FP data by using an integer gather and casting the result. 4257 if (VT.isFloatingPoint()) { 4258 EVT PassThruVT = getPackedSVEVectorVT(VT.getVectorElementCount()); 4259 PassThru = getSVESafeBitCast(PassThruVT, PassThru, DAG); 4260 InputVT = DAG.getValueType(MemVT.changeVectorElementTypeToInteger()); 4261 } 4262 4263 SDVTList VTs = DAG.getVTList(PassThru.getSimpleValueType(), MVT::Other); 4264 4265 if (getGatherScatterIndexIsExtended(Index)) 4266 Index = Index.getOperand(0); 4267 4268 unsigned Opcode = getGatherVecOpcode(IsScaled, IsSigned, IdxNeedsExtend); 4269 selectGatherScatterAddrMode(BasePtr, Index, MemVT, Opcode, 4270 /*isGather=*/true, DAG); 4271 4272 if (ResNeedsSignExtend) 4273 Opcode = getSignExtendedGatherOpcode(Opcode); 4274 4275 SDValue Ops[] = {Chain, Mask, BasePtr, Index, InputVT, PassThru}; 4276 SDValue Gather = DAG.getNode(Opcode, DL, VTs, Ops); 4277 4278 if (VT.isFloatingPoint()) { 4279 SDValue Cast = getSVESafeBitCast(VT, Gather, DAG); 4280 return DAG.getMergeValues({Cast, Gather.getValue(1)}, DL); 4281 } 4282 4283 return Gather; 4284 } 4285 4286 SDValue AArch64TargetLowering::LowerMSCATTER(SDValue Op, 4287 SelectionDAG &DAG) const { 4288 SDLoc DL(Op); 4289 MaskedScatterSDNode *MSC = cast<MaskedScatterSDNode>(Op); 4290 assert(MSC && "Can only custom lower scatter store nodes"); 4291 4292 SDValue Index = MSC->getIndex(); 4293 SDValue Chain = MSC->getChain(); 4294 SDValue StoreVal = MSC->getValue(); 4295 SDValue Mask = MSC->getMask(); 4296 SDValue BasePtr = MSC->getBasePtr(); 4297 4298 ISD::MemIndexType IndexType = MSC->getIndexType(); 4299 bool IsScaled = 4300 IndexType == ISD::SIGNED_SCALED || IndexType == ISD::UNSIGNED_SCALED; 4301 bool IsSigned = 4302 IndexType == ISD::SIGNED_SCALED || IndexType == ISD::SIGNED_UNSCALED; 4303 bool NeedsExtend = 4304 getGatherScatterIndexIsExtended(Index) || 4305 Index.getSimpleValueType().getVectorElementType() == MVT::i32; 4306 4307 EVT VT = StoreVal.getSimpleValueType(); 4308 SDVTList VTs = DAG.getVTList(MVT::Other); 4309 EVT MemVT = MSC->getMemoryVT(); 4310 SDValue InputVT = DAG.getValueType(MemVT); 4311 4312 if (VT.getVectorElementType() == MVT::bf16 && 4313 !static_cast<const AArch64Subtarget &>(DAG.getSubtarget()).hasBF16()) 4314 return SDValue(); 4315 4316 // Handle FP data by casting the data so an integer scatter can be used. 4317 if (VT.isFloatingPoint()) { 4318 EVT StoreValVT = getPackedSVEVectorVT(VT.getVectorElementCount()); 4319 StoreVal = getSVESafeBitCast(StoreValVT, StoreVal, DAG); 4320 InputVT = DAG.getValueType(MemVT.changeVectorElementTypeToInteger()); 4321 } 4322 4323 if (getGatherScatterIndexIsExtended(Index)) 4324 Index = Index.getOperand(0); 4325 4326 unsigned Opcode = getScatterVecOpcode(IsScaled, IsSigned, NeedsExtend); 4327 selectGatherScatterAddrMode(BasePtr, Index, MemVT, Opcode, 4328 /*isGather=*/false, DAG); 4329 4330 SDValue Ops[] = {Chain, StoreVal, Mask, BasePtr, Index, InputVT}; 4331 return DAG.getNode(Opcode, DL, VTs, Ops); 4332 } 4333 4334 // Custom lower trunc store for v4i8 vectors, since it is promoted to v4i16. 4335 static SDValue LowerTruncateVectorStore(SDLoc DL, StoreSDNode *ST, 4336 EVT VT, EVT MemVT, 4337 SelectionDAG &DAG) { 4338 assert(VT.isVector() && "VT should be a vector type"); 4339 assert(MemVT == MVT::v4i8 && VT == MVT::v4i16); 4340 4341 SDValue Value = ST->getValue(); 4342 4343 // It first extend the promoted v4i16 to v8i16, truncate to v8i8, and extract 4344 // the word lane which represent the v4i8 subvector. It optimizes the store 4345 // to: 4346 // 4347 // xtn v0.8b, v0.8h 4348 // str s0, [x0] 4349 4350 SDValue Undef = DAG.getUNDEF(MVT::i16); 4351 SDValue UndefVec = DAG.getBuildVector(MVT::v4i16, DL, 4352 {Undef, Undef, Undef, Undef}); 4353 4354 SDValue TruncExt = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v8i16, 4355 Value, UndefVec); 4356 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, MVT::v8i8, TruncExt); 4357 4358 Trunc = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Trunc); 4359 SDValue ExtractTrunc = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, 4360 Trunc, DAG.getConstant(0, DL, MVT::i64)); 4361 4362 return DAG.getStore(ST->getChain(), DL, ExtractTrunc, 4363 ST->getBasePtr(), ST->getMemOperand()); 4364 } 4365 4366 // Custom lowering for any store, vector or scalar and/or default or with 4367 // a truncate operations. Currently only custom lower truncate operation 4368 // from vector v4i16 to v4i8 or volatile stores of i128. 4369 SDValue AArch64TargetLowering::LowerSTORE(SDValue Op, 4370 SelectionDAG &DAG) const { 4371 SDLoc Dl(Op); 4372 StoreSDNode *StoreNode = cast<StoreSDNode>(Op); 4373 assert (StoreNode && "Can only custom lower store nodes"); 4374 4375 SDValue Value = StoreNode->getValue(); 4376 4377 EVT VT = Value.getValueType(); 4378 EVT MemVT = StoreNode->getMemoryVT(); 4379 4380 if (VT.isVector()) { 4381 if (useSVEForFixedLengthVectorVT(VT)) 4382 return LowerFixedLengthVectorStoreToSVE(Op, DAG); 4383 4384 unsigned AS = StoreNode->getAddressSpace(); 4385 Align Alignment = StoreNode->getAlign(); 4386 if (Alignment < MemVT.getStoreSize() && 4387 !allowsMisalignedMemoryAccesses(MemVT, AS, Alignment, 4388 StoreNode->getMemOperand()->getFlags(), 4389 nullptr)) { 4390 return scalarizeVectorStore(StoreNode, DAG); 4391 } 4392 4393 if (StoreNode->isTruncatingStore()) { 4394 return LowerTruncateVectorStore(Dl, StoreNode, VT, MemVT, DAG); 4395 } 4396 // 256 bit non-temporal stores can be lowered to STNP. Do this as part of 4397 // the custom lowering, as there are no un-paired non-temporal stores and 4398 // legalization will break up 256 bit inputs. 4399 ElementCount EC = MemVT.getVectorElementCount(); 4400 if (StoreNode->isNonTemporal() && MemVT.getSizeInBits() == 256u && 4401 EC.isKnownEven() && 4402 ((MemVT.getScalarSizeInBits() == 8u || 4403 MemVT.getScalarSizeInBits() == 16u || 4404 MemVT.getScalarSizeInBits() == 32u || 4405 MemVT.getScalarSizeInBits() == 64u))) { 4406 SDValue Lo = 4407 DAG.getNode(ISD::EXTRACT_SUBVECTOR, Dl, 4408 MemVT.getHalfNumVectorElementsVT(*DAG.getContext()), 4409 StoreNode->getValue(), DAG.getConstant(0, Dl, MVT::i64)); 4410 SDValue Hi = 4411 DAG.getNode(ISD::EXTRACT_SUBVECTOR, Dl, 4412 MemVT.getHalfNumVectorElementsVT(*DAG.getContext()), 4413 StoreNode->getValue(), 4414 DAG.getConstant(EC.getKnownMinValue() / 2, Dl, MVT::i64)); 4415 SDValue Result = DAG.getMemIntrinsicNode( 4416 AArch64ISD::STNP, Dl, DAG.getVTList(MVT::Other), 4417 {StoreNode->getChain(), Lo, Hi, StoreNode->getBasePtr()}, 4418 StoreNode->getMemoryVT(), StoreNode->getMemOperand()); 4419 return Result; 4420 } 4421 } else if (MemVT == MVT::i128 && StoreNode->isVolatile()) { 4422 assert(StoreNode->getValue()->getValueType(0) == MVT::i128); 4423 SDValue Lo = 4424 DAG.getNode(ISD::EXTRACT_ELEMENT, Dl, MVT::i64, StoreNode->getValue(), 4425 DAG.getConstant(0, Dl, MVT::i64)); 4426 SDValue Hi = 4427 DAG.getNode(ISD::EXTRACT_ELEMENT, Dl, MVT::i64, StoreNode->getValue(), 4428 DAG.getConstant(1, Dl, MVT::i64)); 4429 SDValue Result = DAG.getMemIntrinsicNode( 4430 AArch64ISD::STP, Dl, DAG.getVTList(MVT::Other), 4431 {StoreNode->getChain(), Lo, Hi, StoreNode->getBasePtr()}, 4432 StoreNode->getMemoryVT(), StoreNode->getMemOperand()); 4433 return Result; 4434 } 4435 4436 return SDValue(); 4437 } 4438 4439 // Generate SUBS and CSEL for integer abs. 4440 SDValue AArch64TargetLowering::LowerABS(SDValue Op, SelectionDAG &DAG) const { 4441 MVT VT = Op.getSimpleValueType(); 4442 4443 if (VT.isVector()) 4444 return LowerToPredicatedOp(Op, DAG, AArch64ISD::ABS_MERGE_PASSTHRU); 4445 4446 SDLoc DL(Op); 4447 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), 4448 Op.getOperand(0)); 4449 // Generate SUBS & CSEL. 4450 SDValue Cmp = 4451 DAG.getNode(AArch64ISD::SUBS, DL, DAG.getVTList(VT, MVT::i32), 4452 Op.getOperand(0), DAG.getConstant(0, DL, VT)); 4453 return DAG.getNode(AArch64ISD::CSEL, DL, VT, Op.getOperand(0), Neg, 4454 DAG.getConstant(AArch64CC::PL, DL, MVT::i32), 4455 Cmp.getValue(1)); 4456 } 4457 4458 SDValue AArch64TargetLowering::LowerOperation(SDValue Op, 4459 SelectionDAG &DAG) const { 4460 LLVM_DEBUG(dbgs() << "Custom lowering: "); 4461 LLVM_DEBUG(Op.dump()); 4462 4463 switch (Op.getOpcode()) { 4464 default: 4465 llvm_unreachable("unimplemented operand"); 4466 return SDValue(); 4467 case ISD::BITCAST: 4468 return LowerBITCAST(Op, DAG); 4469 case ISD::GlobalAddress: 4470 return LowerGlobalAddress(Op, DAG); 4471 case ISD::GlobalTLSAddress: 4472 return LowerGlobalTLSAddress(Op, DAG); 4473 case ISD::SETCC: 4474 case ISD::STRICT_FSETCC: 4475 case ISD::STRICT_FSETCCS: 4476 return LowerSETCC(Op, DAG); 4477 case ISD::BR_CC: 4478 return LowerBR_CC(Op, DAG); 4479 case ISD::SELECT: 4480 return LowerSELECT(Op, DAG); 4481 case ISD::SELECT_CC: 4482 return LowerSELECT_CC(Op, DAG); 4483 case ISD::JumpTable: 4484 return LowerJumpTable(Op, DAG); 4485 case ISD::BR_JT: 4486 return LowerBR_JT(Op, DAG); 4487 case ISD::ConstantPool: 4488 return LowerConstantPool(Op, DAG); 4489 case ISD::BlockAddress: 4490 return LowerBlockAddress(Op, DAG); 4491 case ISD::VASTART: 4492 return LowerVASTART(Op, DAG); 4493 case ISD::VACOPY: 4494 return LowerVACOPY(Op, DAG); 4495 case ISD::VAARG: 4496 return LowerVAARG(Op, DAG); 4497 case ISD::ADDC: 4498 case ISD::ADDE: 4499 case ISD::SUBC: 4500 case ISD::SUBE: 4501 return LowerADDC_ADDE_SUBC_SUBE(Op, DAG); 4502 case ISD::SADDO: 4503 case ISD::UADDO: 4504 case ISD::SSUBO: 4505 case ISD::USUBO: 4506 case ISD::SMULO: 4507 case ISD::UMULO: 4508 return LowerXALUO(Op, DAG); 4509 case ISD::FADD: 4510 return LowerToPredicatedOp(Op, DAG, AArch64ISD::FADD_PRED); 4511 case ISD::FSUB: 4512 return LowerToPredicatedOp(Op, DAG, AArch64ISD::FSUB_PRED); 4513 case ISD::FMUL: 4514 return LowerToPredicatedOp(Op, DAG, AArch64ISD::FMUL_PRED); 4515 case ISD::FMA: 4516 return LowerToPredicatedOp(Op, DAG, AArch64ISD::FMA_PRED); 4517 case ISD::FDIV: 4518 return LowerToPredicatedOp(Op, DAG, AArch64ISD::FDIV_PRED); 4519 case ISD::FNEG: 4520 return LowerToPredicatedOp(Op, DAG, AArch64ISD::FNEG_MERGE_PASSTHRU); 4521 case ISD::FCEIL: 4522 return LowerToPredicatedOp(Op, DAG, AArch64ISD::FCEIL_MERGE_PASSTHRU); 4523 case ISD::FFLOOR: 4524 return LowerToPredicatedOp(Op, DAG, AArch64ISD::FFLOOR_MERGE_PASSTHRU); 4525 case ISD::FNEARBYINT: 4526 return LowerToPredicatedOp(Op, DAG, AArch64ISD::FNEARBYINT_MERGE_PASSTHRU); 4527 case ISD::FRINT: 4528 return LowerToPredicatedOp(Op, DAG, AArch64ISD::FRINT_MERGE_PASSTHRU); 4529 case ISD::FROUND: 4530 return LowerToPredicatedOp(Op, DAG, AArch64ISD::FROUND_MERGE_PASSTHRU); 4531 case ISD::FROUNDEVEN: 4532 return LowerToPredicatedOp(Op, DAG, AArch64ISD::FROUNDEVEN_MERGE_PASSTHRU); 4533 case ISD::FTRUNC: 4534 return LowerToPredicatedOp(Op, DAG, AArch64ISD::FTRUNC_MERGE_PASSTHRU); 4535 case ISD::FSQRT: 4536 return LowerToPredicatedOp(Op, DAG, AArch64ISD::FSQRT_MERGE_PASSTHRU); 4537 case ISD::FABS: 4538 return LowerToPredicatedOp(Op, DAG, AArch64ISD::FABS_MERGE_PASSTHRU); 4539 case ISD::FP_ROUND: 4540 case ISD::STRICT_FP_ROUND: 4541 return LowerFP_ROUND(Op, DAG); 4542 case ISD::FP_EXTEND: 4543 return LowerFP_EXTEND(Op, DAG); 4544 case ISD::FRAMEADDR: 4545 return LowerFRAMEADDR(Op, DAG); 4546 case ISD::SPONENTRY: 4547 return LowerSPONENTRY(Op, DAG); 4548 case ISD::RETURNADDR: 4549 return LowerRETURNADDR(Op, DAG); 4550 case ISD::ADDROFRETURNADDR: 4551 return LowerADDROFRETURNADDR(Op, DAG); 4552 case ISD::CONCAT_VECTORS: 4553 return LowerCONCAT_VECTORS(Op, DAG); 4554 case ISD::INSERT_VECTOR_ELT: 4555 return LowerINSERT_VECTOR_ELT(Op, DAG); 4556 case ISD::EXTRACT_VECTOR_ELT: 4557 return LowerEXTRACT_VECTOR_ELT(Op, DAG); 4558 case ISD::BUILD_VECTOR: 4559 return LowerBUILD_VECTOR(Op, DAG); 4560 case ISD::VECTOR_SHUFFLE: 4561 return LowerVECTOR_SHUFFLE(Op, DAG); 4562 case ISD::SPLAT_VECTOR: 4563 return LowerSPLAT_VECTOR(Op, DAG); 4564 case ISD::EXTRACT_SUBVECTOR: 4565 return LowerEXTRACT_SUBVECTOR(Op, DAG); 4566 case ISD::INSERT_SUBVECTOR: 4567 return LowerINSERT_SUBVECTOR(Op, DAG); 4568 case ISD::SDIV: 4569 case ISD::UDIV: 4570 return LowerDIV(Op, DAG); 4571 case ISD::SMIN: 4572 return LowerToPredicatedOp(Op, DAG, AArch64ISD::SMIN_PRED, 4573 /*OverrideNEON=*/true); 4574 case ISD::UMIN: 4575 return LowerToPredicatedOp(Op, DAG, AArch64ISD::UMIN_PRED, 4576 /*OverrideNEON=*/true); 4577 case ISD::SMAX: 4578 return LowerToPredicatedOp(Op, DAG, AArch64ISD::SMAX_PRED, 4579 /*OverrideNEON=*/true); 4580 case ISD::UMAX: 4581 return LowerToPredicatedOp(Op, DAG, AArch64ISD::UMAX_PRED, 4582 /*OverrideNEON=*/true); 4583 case ISD::SRA: 4584 case ISD::SRL: 4585 case ISD::SHL: 4586 return LowerVectorSRA_SRL_SHL(Op, DAG); 4587 case ISD::SHL_PARTS: 4588 case ISD::SRL_PARTS: 4589 case ISD::SRA_PARTS: 4590 return LowerShiftParts(Op, DAG); 4591 case ISD::CTPOP: 4592 return LowerCTPOP(Op, DAG); 4593 case ISD::FCOPYSIGN: 4594 return LowerFCOPYSIGN(Op, DAG); 4595 case ISD::OR: 4596 return LowerVectorOR(Op, DAG); 4597 case ISD::XOR: 4598 return LowerXOR(Op, DAG); 4599 case ISD::PREFETCH: 4600 return LowerPREFETCH(Op, DAG); 4601 case ISD::SINT_TO_FP: 4602 case ISD::UINT_TO_FP: 4603 case ISD::STRICT_SINT_TO_FP: 4604 case ISD::STRICT_UINT_TO_FP: 4605 return LowerINT_TO_FP(Op, DAG); 4606 case ISD::FP_TO_SINT: 4607 case ISD::FP_TO_UINT: 4608 case ISD::STRICT_FP_TO_SINT: 4609 case ISD::STRICT_FP_TO_UINT: 4610 return LowerFP_TO_INT(Op, DAG); 4611 case ISD::FP_TO_SINT_SAT: 4612 case ISD::FP_TO_UINT_SAT: 4613 return LowerFP_TO_INT_SAT(Op, DAG); 4614 case ISD::FSINCOS: 4615 return LowerFSINCOS(Op, DAG); 4616 case ISD::FLT_ROUNDS_: 4617 return LowerFLT_ROUNDS_(Op, DAG); 4618 case ISD::SET_ROUNDING: 4619 return LowerSET_ROUNDING(Op, DAG); 4620 case ISD::MUL: 4621 return LowerMUL(Op, DAG); 4622 case ISD::MULHS: 4623 return LowerToPredicatedOp(Op, DAG, AArch64ISD::MULHS_PRED, 4624 /*OverrideNEON=*/true); 4625 case ISD::MULHU: 4626 return LowerToPredicatedOp(Op, DAG, AArch64ISD::MULHU_PRED, 4627 /*OverrideNEON=*/true); 4628 case ISD::INTRINSIC_WO_CHAIN: 4629 return LowerINTRINSIC_WO_CHAIN(Op, DAG); 4630 case ISD::STORE: 4631 return LowerSTORE(Op, DAG); 4632 case ISD::MSTORE: 4633 return LowerFixedLengthVectorMStoreToSVE(Op, DAG); 4634 case ISD::MGATHER: 4635 return LowerMGATHER(Op, DAG); 4636 case ISD::MSCATTER: 4637 return LowerMSCATTER(Op, DAG); 4638 case ISD::VECREDUCE_SEQ_FADD: 4639 return LowerVECREDUCE_SEQ_FADD(Op, DAG); 4640 case ISD::VECREDUCE_ADD: 4641 case ISD::VECREDUCE_AND: 4642 case ISD::VECREDUCE_OR: 4643 case ISD::VECREDUCE_XOR: 4644 case ISD::VECREDUCE_SMAX: 4645 case ISD::VECREDUCE_SMIN: 4646 case ISD::VECREDUCE_UMAX: 4647 case ISD::VECREDUCE_UMIN: 4648 case ISD::VECREDUCE_FADD: 4649 case ISD::VECREDUCE_FMAX: 4650 case ISD::VECREDUCE_FMIN: 4651 return LowerVECREDUCE(Op, DAG); 4652 case ISD::ATOMIC_LOAD_SUB: 4653 return LowerATOMIC_LOAD_SUB(Op, DAG); 4654 case ISD::ATOMIC_LOAD_AND: 4655 return LowerATOMIC_LOAD_AND(Op, DAG); 4656 case ISD::DYNAMIC_STACKALLOC: 4657 return LowerDYNAMIC_STACKALLOC(Op, DAG); 4658 case ISD::VSCALE: 4659 return LowerVSCALE(Op, DAG); 4660 case ISD::ANY_EXTEND: 4661 case ISD::SIGN_EXTEND: 4662 case ISD::ZERO_EXTEND: 4663 return LowerFixedLengthVectorIntExtendToSVE(Op, DAG); 4664 case ISD::SIGN_EXTEND_INREG: { 4665 // Only custom lower when ExtraVT has a legal byte based element type. 4666 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 4667 EVT ExtraEltVT = ExtraVT.getVectorElementType(); 4668 if ((ExtraEltVT != MVT::i8) && (ExtraEltVT != MVT::i16) && 4669 (ExtraEltVT != MVT::i32) && (ExtraEltVT != MVT::i64)) 4670 return SDValue(); 4671 4672 return LowerToPredicatedOp(Op, DAG, 4673 AArch64ISD::SIGN_EXTEND_INREG_MERGE_PASSTHRU); 4674 } 4675 case ISD::TRUNCATE: 4676 return LowerTRUNCATE(Op, DAG); 4677 case ISD::MLOAD: 4678 return LowerFixedLengthVectorMLoadToSVE(Op, DAG); 4679 case ISD::LOAD: 4680 if (useSVEForFixedLengthVectorVT(Op.getValueType())) 4681 return LowerFixedLengthVectorLoadToSVE(Op, DAG); 4682 llvm_unreachable("Unexpected request to lower ISD::LOAD"); 4683 case ISD::ADD: 4684 return LowerToPredicatedOp(Op, DAG, AArch64ISD::ADD_PRED); 4685 case ISD::AND: 4686 return LowerToScalableOp(Op, DAG); 4687 case ISD::SUB: 4688 return LowerToPredicatedOp(Op, DAG, AArch64ISD::SUB_PRED); 4689 case ISD::FMAXIMUM: 4690 return LowerToPredicatedOp(Op, DAG, AArch64ISD::FMAX_PRED); 4691 case ISD::FMAXNUM: 4692 return LowerToPredicatedOp(Op, DAG, AArch64ISD::FMAXNM_PRED); 4693 case ISD::FMINIMUM: 4694 return LowerToPredicatedOp(Op, DAG, AArch64ISD::FMIN_PRED); 4695 case ISD::FMINNUM: 4696 return LowerToPredicatedOp(Op, DAG, AArch64ISD::FMINNM_PRED); 4697 case ISD::VSELECT: 4698 return LowerFixedLengthVectorSelectToSVE(Op, DAG); 4699 case ISD::ABS: 4700 return LowerABS(Op, DAG); 4701 case ISD::BITREVERSE: 4702 return LowerToPredicatedOp(Op, DAG, AArch64ISD::BITREVERSE_MERGE_PASSTHRU, 4703 /*OverrideNEON=*/true); 4704 case ISD::BSWAP: 4705 return LowerToPredicatedOp(Op, DAG, AArch64ISD::BSWAP_MERGE_PASSTHRU); 4706 case ISD::CTLZ: 4707 return LowerToPredicatedOp(Op, DAG, AArch64ISD::CTLZ_MERGE_PASSTHRU, 4708 /*OverrideNEON=*/true); 4709 case ISD::CTTZ: 4710 return LowerCTTZ(Op, DAG); 4711 } 4712 } 4713 4714 bool AArch64TargetLowering::mergeStoresAfterLegalization(EVT VT) const { 4715 return !Subtarget->useSVEForFixedLengthVectors(); 4716 } 4717 4718 bool AArch64TargetLowering::useSVEForFixedLengthVectorVT( 4719 EVT VT, bool OverrideNEON) const { 4720 if (!Subtarget->useSVEForFixedLengthVectors()) 4721 return false; 4722 4723 if (!VT.isFixedLengthVector()) 4724 return false; 4725 4726 // Don't use SVE for vectors we cannot scalarize if required. 4727 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) { 4728 // Fixed length predicates should be promoted to i8. 4729 // NOTE: This is consistent with how NEON (and thus 64/128bit vectors) work. 4730 case MVT::i1: 4731 default: 4732 return false; 4733 case MVT::i8: 4734 case MVT::i16: 4735 case MVT::i32: 4736 case MVT::i64: 4737 case MVT::f16: 4738 case MVT::f32: 4739 case MVT::f64: 4740 break; 4741 } 4742 4743 // All SVE implementations support NEON sized vectors. 4744 if (OverrideNEON && (VT.is128BitVector() || VT.is64BitVector())) 4745 return true; 4746 4747 // Ensure NEON MVTs only belong to a single register class. 4748 if (VT.getFixedSizeInBits() <= 128) 4749 return false; 4750 4751 // Don't use SVE for types that don't fit. 4752 if (VT.getFixedSizeInBits() > Subtarget->getMinSVEVectorSizeInBits()) 4753 return false; 4754 4755 // TODO: Perhaps an artificial restriction, but worth having whilst getting 4756 // the base fixed length SVE support in place. 4757 if (!VT.isPow2VectorType()) 4758 return false; 4759 4760 return true; 4761 } 4762 4763 //===----------------------------------------------------------------------===// 4764 // Calling Convention Implementation 4765 //===----------------------------------------------------------------------===// 4766 4767 /// Selects the correct CCAssignFn for a given CallingConvention value. 4768 CCAssignFn *AArch64TargetLowering::CCAssignFnForCall(CallingConv::ID CC, 4769 bool IsVarArg) const { 4770 switch (CC) { 4771 default: 4772 report_fatal_error("Unsupported calling convention."); 4773 case CallingConv::WebKit_JS: 4774 return CC_AArch64_WebKit_JS; 4775 case CallingConv::GHC: 4776 return CC_AArch64_GHC; 4777 case CallingConv::C: 4778 case CallingConv::Fast: 4779 case CallingConv::PreserveMost: 4780 case CallingConv::CXX_FAST_TLS: 4781 case CallingConv::Swift: 4782 case CallingConv::SwiftTail: 4783 case CallingConv::Tail: 4784 if (Subtarget->isTargetWindows() && IsVarArg) 4785 return CC_AArch64_Win64_VarArg; 4786 if (!Subtarget->isTargetDarwin()) 4787 return CC_AArch64_AAPCS; 4788 if (!IsVarArg) 4789 return CC_AArch64_DarwinPCS; 4790 return Subtarget->isTargetILP32() ? CC_AArch64_DarwinPCS_ILP32_VarArg 4791 : CC_AArch64_DarwinPCS_VarArg; 4792 case CallingConv::Win64: 4793 return IsVarArg ? CC_AArch64_Win64_VarArg : CC_AArch64_AAPCS; 4794 case CallingConv::CFGuard_Check: 4795 return CC_AArch64_Win64_CFGuard_Check; 4796 case CallingConv::AArch64_VectorCall: 4797 case CallingConv::AArch64_SVE_VectorCall: 4798 return CC_AArch64_AAPCS; 4799 } 4800 } 4801 4802 CCAssignFn * 4803 AArch64TargetLowering::CCAssignFnForReturn(CallingConv::ID CC) const { 4804 return CC == CallingConv::WebKit_JS ? RetCC_AArch64_WebKit_JS 4805 : RetCC_AArch64_AAPCS; 4806 } 4807 4808 SDValue AArch64TargetLowering::LowerFormalArguments( 4809 SDValue Chain, CallingConv::ID CallConv, bool isVarArg, 4810 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, 4811 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { 4812 MachineFunction &MF = DAG.getMachineFunction(); 4813 MachineFrameInfo &MFI = MF.getFrameInfo(); 4814 bool IsWin64 = Subtarget->isCallingConvWin64(MF.getFunction().getCallingConv()); 4815 4816 // Assign locations to all of the incoming arguments. 4817 SmallVector<CCValAssign, 16> ArgLocs; 4818 DenseMap<unsigned, SDValue> CopiedRegs; 4819 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs, 4820 *DAG.getContext()); 4821 4822 // At this point, Ins[].VT may already be promoted to i32. To correctly 4823 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and 4824 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT. 4825 // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here 4826 // we use a special version of AnalyzeFormalArguments to pass in ValVT and 4827 // LocVT. 4828 unsigned NumArgs = Ins.size(); 4829 Function::const_arg_iterator CurOrigArg = MF.getFunction().arg_begin(); 4830 unsigned CurArgIdx = 0; 4831 for (unsigned i = 0; i != NumArgs; ++i) { 4832 MVT ValVT = Ins[i].VT; 4833 if (Ins[i].isOrigArg()) { 4834 std::advance(CurOrigArg, Ins[i].getOrigArgIndex() - CurArgIdx); 4835 CurArgIdx = Ins[i].getOrigArgIndex(); 4836 4837 // Get type of the original argument. 4838 EVT ActualVT = getValueType(DAG.getDataLayout(), CurOrigArg->getType(), 4839 /*AllowUnknown*/ true); 4840 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other; 4841 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16. 4842 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8) 4843 ValVT = MVT::i8; 4844 else if (ActualMVT == MVT::i16) 4845 ValVT = MVT::i16; 4846 } 4847 bool UseVarArgCC = false; 4848 if (IsWin64) 4849 UseVarArgCC = isVarArg; 4850 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, UseVarArgCC); 4851 bool Res = 4852 AssignFn(i, ValVT, ValVT, CCValAssign::Full, Ins[i].Flags, CCInfo); 4853 assert(!Res && "Call operand has unhandled type"); 4854 (void)Res; 4855 } 4856 SmallVector<SDValue, 16> ArgValues; 4857 unsigned ExtraArgLocs = 0; 4858 for (unsigned i = 0, e = Ins.size(); i != e; ++i) { 4859 CCValAssign &VA = ArgLocs[i - ExtraArgLocs]; 4860 4861 if (Ins[i].Flags.isByVal()) { 4862 // Byval is used for HFAs in the PCS, but the system should work in a 4863 // non-compliant manner for larger structs. 4864 EVT PtrVT = getPointerTy(DAG.getDataLayout()); 4865 int Size = Ins[i].Flags.getByValSize(); 4866 unsigned NumRegs = (Size + 7) / 8; 4867 4868 // FIXME: This works on big-endian for composite byvals, which are the common 4869 // case. It should also work for fundamental types too. 4870 unsigned FrameIdx = 4871 MFI.CreateFixedObject(8 * NumRegs, VA.getLocMemOffset(), false); 4872 SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrVT); 4873 InVals.push_back(FrameIdxN); 4874 4875 continue; 4876 } 4877 4878 if (Ins[i].Flags.isSwiftAsync()) 4879 MF.getInfo<AArch64FunctionInfo>()->setHasSwiftAsyncContext(true); 4880 4881 SDValue ArgValue; 4882 if (VA.isRegLoc()) { 4883 // Arguments stored in registers. 4884 EVT RegVT = VA.getLocVT(); 4885 const TargetRegisterClass *RC; 4886 4887 if (RegVT == MVT::i32) 4888 RC = &AArch64::GPR32RegClass; 4889 else if (RegVT == MVT::i64) 4890 RC = &AArch64::GPR64RegClass; 4891 else if (RegVT == MVT::f16 || RegVT == MVT::bf16) 4892 RC = &AArch64::FPR16RegClass; 4893 else if (RegVT == MVT::f32) 4894 RC = &AArch64::FPR32RegClass; 4895 else if (RegVT == MVT::f64 || RegVT.is64BitVector()) 4896 RC = &AArch64::FPR64RegClass; 4897 else if (RegVT == MVT::f128 || RegVT.is128BitVector()) 4898 RC = &AArch64::FPR128RegClass; 4899 else if (RegVT.isScalableVector() && 4900 RegVT.getVectorElementType() == MVT::i1) 4901 RC = &AArch64::PPRRegClass; 4902 else if (RegVT.isScalableVector()) 4903 RC = &AArch64::ZPRRegClass; 4904 else 4905 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering"); 4906 4907 // Transform the arguments in physical registers into virtual ones. 4908 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC); 4909 ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT); 4910 4911 // If this is an 8, 16 or 32-bit value, it is really passed promoted 4912 // to 64 bits. Insert an assert[sz]ext to capture this, then 4913 // truncate to the right size. 4914 switch (VA.getLocInfo()) { 4915 default: 4916 llvm_unreachable("Unknown loc info!"); 4917 case CCValAssign::Full: 4918 break; 4919 case CCValAssign::Indirect: 4920 assert(VA.getValVT().isScalableVector() && 4921 "Only scalable vectors can be passed indirectly"); 4922 break; 4923 case CCValAssign::BCvt: 4924 ArgValue = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), ArgValue); 4925 break; 4926 case CCValAssign::AExt: 4927 case CCValAssign::SExt: 4928 case CCValAssign::ZExt: 4929 break; 4930 case CCValAssign::AExtUpper: 4931 ArgValue = DAG.getNode(ISD::SRL, DL, RegVT, ArgValue, 4932 DAG.getConstant(32, DL, RegVT)); 4933 ArgValue = DAG.getZExtOrTrunc(ArgValue, DL, VA.getValVT()); 4934 break; 4935 } 4936 } else { // VA.isRegLoc() 4937 assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem"); 4938 unsigned ArgOffset = VA.getLocMemOffset(); 4939 unsigned ArgSize = (VA.getLocInfo() == CCValAssign::Indirect 4940 ? VA.getLocVT().getSizeInBits() 4941 : VA.getValVT().getSizeInBits()) / 8; 4942 4943 uint32_t BEAlign = 0; 4944 if (!Subtarget->isLittleEndian() && ArgSize < 8 && 4945 !Ins[i].Flags.isInConsecutiveRegs()) 4946 BEAlign = 8 - ArgSize; 4947 4948 int FI = MFI.CreateFixedObject(ArgSize, ArgOffset + BEAlign, true); 4949 4950 // Create load nodes to retrieve arguments from the stack. 4951 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout())); 4952 4953 // For NON_EXTLOAD, generic code in getLoad assert(ValVT == MemVT) 4954 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD; 4955 MVT MemVT = VA.getValVT(); 4956 4957 switch (VA.getLocInfo()) { 4958 default: 4959 break; 4960 case CCValAssign::Trunc: 4961 case CCValAssign::BCvt: 4962 MemVT = VA.getLocVT(); 4963 break; 4964 case CCValAssign::Indirect: 4965 assert(VA.getValVT().isScalableVector() && 4966 "Only scalable vectors can be passed indirectly"); 4967 MemVT = VA.getLocVT(); 4968 break; 4969 case CCValAssign::SExt: 4970 ExtType = ISD::SEXTLOAD; 4971 break; 4972 case CCValAssign::ZExt: 4973 ExtType = ISD::ZEXTLOAD; 4974 break; 4975 case CCValAssign::AExt: 4976 ExtType = ISD::EXTLOAD; 4977 break; 4978 } 4979 4980 ArgValue = DAG.getExtLoad( 4981 ExtType, DL, VA.getLocVT(), Chain, FIN, 4982 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), 4983 MemVT); 4984 4985 } 4986 4987 if (VA.getLocInfo() == CCValAssign::Indirect) { 4988 assert(VA.getValVT().isScalableVector() && 4989 "Only scalable vectors can be passed indirectly"); 4990 4991 uint64_t PartSize = VA.getValVT().getStoreSize().getKnownMinSize(); 4992 unsigned NumParts = 1; 4993 if (Ins[i].Flags.isInConsecutiveRegs()) { 4994 assert(!Ins[i].Flags.isInConsecutiveRegsLast()); 4995 while (!Ins[i + NumParts - 1].Flags.isInConsecutiveRegsLast()) 4996 ++NumParts; 4997 } 4998 4999 MVT PartLoad = VA.getValVT(); 5000 SDValue Ptr = ArgValue; 5001 5002 // Ensure we generate all loads for each tuple part, whilst updating the 5003 // pointer after each load correctly using vscale. 5004 while (NumParts > 0) { 5005 ArgValue = DAG.getLoad(PartLoad, DL, Chain, Ptr, MachinePointerInfo()); 5006 InVals.push_back(ArgValue); 5007 NumParts--; 5008 if (NumParts > 0) { 5009 SDValue BytesIncrement = DAG.getVScale( 5010 DL, Ptr.getValueType(), 5011 APInt(Ptr.getValueSizeInBits().getFixedSize(), PartSize)); 5012 SDNodeFlags Flags; 5013 Flags.setNoUnsignedWrap(true); 5014 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr, 5015 BytesIncrement, Flags); 5016 ExtraArgLocs++; 5017 i++; 5018 } 5019 } 5020 } else { 5021 if (Subtarget->isTargetILP32() && Ins[i].Flags.isPointer()) 5022 ArgValue = DAG.getNode(ISD::AssertZext, DL, ArgValue.getValueType(), 5023 ArgValue, DAG.getValueType(MVT::i32)); 5024 InVals.push_back(ArgValue); 5025 } 5026 } 5027 assert((ArgLocs.size() + ExtraArgLocs) == Ins.size()); 5028 5029 // varargs 5030 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>(); 5031 if (isVarArg) { 5032 if (!Subtarget->isTargetDarwin() || IsWin64) { 5033 // The AAPCS variadic function ABI is identical to the non-variadic 5034 // one. As a result there may be more arguments in registers and we should 5035 // save them for future reference. 5036 // Win64 variadic functions also pass arguments in registers, but all float 5037 // arguments are passed in integer registers. 5038 saveVarArgRegisters(CCInfo, DAG, DL, Chain); 5039 } 5040 5041 // This will point to the next argument passed via stack. 5042 unsigned StackOffset = CCInfo.getNextStackOffset(); 5043 // We currently pass all varargs at 8-byte alignment, or 4 for ILP32 5044 StackOffset = alignTo(StackOffset, Subtarget->isTargetILP32() ? 4 : 8); 5045 FuncInfo->setVarArgsStackIndex(MFI.CreateFixedObject(4, StackOffset, true)); 5046 5047 if (MFI.hasMustTailInVarArgFunc()) { 5048 SmallVector<MVT, 2> RegParmTypes; 5049 RegParmTypes.push_back(MVT::i64); 5050 RegParmTypes.push_back(MVT::f128); 5051 // Compute the set of forwarded registers. The rest are scratch. 5052 SmallVectorImpl<ForwardedRegister> &Forwards = 5053 FuncInfo->getForwardedMustTailRegParms(); 5054 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, 5055 CC_AArch64_AAPCS); 5056 5057 // Conservatively forward X8, since it might be used for aggregate return. 5058 if (!CCInfo.isAllocated(AArch64::X8)) { 5059 unsigned X8VReg = MF.addLiveIn(AArch64::X8, &AArch64::GPR64RegClass); 5060 Forwards.push_back(ForwardedRegister(X8VReg, AArch64::X8, MVT::i64)); 5061 } 5062 } 5063 } 5064 5065 // On Windows, InReg pointers must be returned, so record the pointer in a 5066 // virtual register at the start of the function so it can be returned in the 5067 // epilogue. 5068 if (IsWin64) { 5069 for (unsigned I = 0, E = Ins.size(); I != E; ++I) { 5070 if (Ins[I].Flags.isInReg()) { 5071 assert(!FuncInfo->getSRetReturnReg()); 5072 5073 MVT PtrTy = getPointerTy(DAG.getDataLayout()); 5074 Register Reg = 5075 MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy)); 5076 FuncInfo->setSRetReturnReg(Reg); 5077 5078 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), DL, Reg, InVals[I]); 5079 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Copy, Chain); 5080 break; 5081 } 5082 } 5083 } 5084 5085 unsigned StackArgSize = CCInfo.getNextStackOffset(); 5086 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt; 5087 if (DoesCalleeRestoreStack(CallConv, TailCallOpt)) { 5088 // This is a non-standard ABI so by fiat I say we're allowed to make full 5089 // use of the stack area to be popped, which must be aligned to 16 bytes in 5090 // any case: 5091 StackArgSize = alignTo(StackArgSize, 16); 5092 5093 // If we're expected to restore the stack (e.g. fastcc) then we'll be adding 5094 // a multiple of 16. 5095 FuncInfo->setArgumentStackToRestore(StackArgSize); 5096 5097 // This realignment carries over to the available bytes below. Our own 5098 // callers will guarantee the space is free by giving an aligned value to 5099 // CALLSEQ_START. 5100 } 5101 // Even if we're not expected to free up the space, it's useful to know how 5102 // much is there while considering tail calls (because we can reuse it). 5103 FuncInfo->setBytesInStackArgArea(StackArgSize); 5104 5105 if (Subtarget->hasCustomCallingConv()) 5106 Subtarget->getRegisterInfo()->UpdateCustomCalleeSavedRegs(MF); 5107 5108 return Chain; 5109 } 5110 5111 void AArch64TargetLowering::saveVarArgRegisters(CCState &CCInfo, 5112 SelectionDAG &DAG, 5113 const SDLoc &DL, 5114 SDValue &Chain) const { 5115 MachineFunction &MF = DAG.getMachineFunction(); 5116 MachineFrameInfo &MFI = MF.getFrameInfo(); 5117 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>(); 5118 auto PtrVT = getPointerTy(DAG.getDataLayout()); 5119 bool IsWin64 = Subtarget->isCallingConvWin64(MF.getFunction().getCallingConv()); 5120 5121 SmallVector<SDValue, 8> MemOps; 5122 5123 static const MCPhysReg GPRArgRegs[] = { AArch64::X0, AArch64::X1, AArch64::X2, 5124 AArch64::X3, AArch64::X4, AArch64::X5, 5125 AArch64::X6, AArch64::X7 }; 5126 static const unsigned NumGPRArgRegs = array_lengthof(GPRArgRegs); 5127 unsigned FirstVariadicGPR = CCInfo.getFirstUnallocated(GPRArgRegs); 5128 5129 unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR); 5130 int GPRIdx = 0; 5131 if (GPRSaveSize != 0) { 5132 if (IsWin64) { 5133 GPRIdx = MFI.CreateFixedObject(GPRSaveSize, -(int)GPRSaveSize, false); 5134 if (GPRSaveSize & 15) 5135 // The extra size here, if triggered, will always be 8. 5136 MFI.CreateFixedObject(16 - (GPRSaveSize & 15), -(int)alignTo(GPRSaveSize, 16), false); 5137 } else 5138 GPRIdx = MFI.CreateStackObject(GPRSaveSize, Align(8), false); 5139 5140 SDValue FIN = DAG.getFrameIndex(GPRIdx, PtrVT); 5141 5142 for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) { 5143 unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &AArch64::GPR64RegClass); 5144 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64); 5145 SDValue Store = DAG.getStore( 5146 Val.getValue(1), DL, Val, FIN, 5147 IsWin64 5148 ? MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), 5149 GPRIdx, 5150 (i - FirstVariadicGPR) * 8) 5151 : MachinePointerInfo::getStack(DAG.getMachineFunction(), i * 8)); 5152 MemOps.push_back(Store); 5153 FIN = 5154 DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getConstant(8, DL, PtrVT)); 5155 } 5156 } 5157 FuncInfo->setVarArgsGPRIndex(GPRIdx); 5158 FuncInfo->setVarArgsGPRSize(GPRSaveSize); 5159 5160 if (Subtarget->hasFPARMv8() && !IsWin64) { 5161 static const MCPhysReg FPRArgRegs[] = { 5162 AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3, 5163 AArch64::Q4, AArch64::Q5, AArch64::Q6, AArch64::Q7}; 5164 static const unsigned NumFPRArgRegs = array_lengthof(FPRArgRegs); 5165 unsigned FirstVariadicFPR = CCInfo.getFirstUnallocated(FPRArgRegs); 5166 5167 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR); 5168 int FPRIdx = 0; 5169 if (FPRSaveSize != 0) { 5170 FPRIdx = MFI.CreateStackObject(FPRSaveSize, Align(16), false); 5171 5172 SDValue FIN = DAG.getFrameIndex(FPRIdx, PtrVT); 5173 5174 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) { 5175 unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &AArch64::FPR128RegClass); 5176 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f128); 5177 5178 SDValue Store = DAG.getStore( 5179 Val.getValue(1), DL, Val, FIN, 5180 MachinePointerInfo::getStack(DAG.getMachineFunction(), i * 16)); 5181 MemOps.push_back(Store); 5182 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, 5183 DAG.getConstant(16, DL, PtrVT)); 5184 } 5185 } 5186 FuncInfo->setVarArgsFPRIndex(FPRIdx); 5187 FuncInfo->setVarArgsFPRSize(FPRSaveSize); 5188 } 5189 5190 if (!MemOps.empty()) { 5191 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps); 5192 } 5193 } 5194 5195 /// LowerCallResult - Lower the result values of a call into the 5196 /// appropriate copies out of appropriate physical registers. 5197 SDValue AArch64TargetLowering::LowerCallResult( 5198 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg, 5199 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, 5200 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, bool isThisReturn, 5201 SDValue ThisVal) const { 5202 CCAssignFn *RetCC = CCAssignFnForReturn(CallConv); 5203 // Assign locations to each value returned by this call. 5204 SmallVector<CCValAssign, 16> RVLocs; 5205 DenseMap<unsigned, SDValue> CopiedRegs; 5206 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs, 5207 *DAG.getContext()); 5208 CCInfo.AnalyzeCallResult(Ins, RetCC); 5209 5210 // Copy all of the result registers out of their specified physreg. 5211 for (unsigned i = 0; i != RVLocs.size(); ++i) { 5212 CCValAssign VA = RVLocs[i]; 5213 5214 // Pass 'this' value directly from the argument to return value, to avoid 5215 // reg unit interference 5216 if (i == 0 && isThisReturn) { 5217 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i64 && 5218 "unexpected return calling convention register assignment"); 5219 InVals.push_back(ThisVal); 5220 continue; 5221 } 5222 5223 // Avoid copying a physreg twice since RegAllocFast is incompetent and only 5224 // allows one use of a physreg per block. 5225 SDValue Val = CopiedRegs.lookup(VA.getLocReg()); 5226 if (!Val) { 5227 Val = 5228 DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InFlag); 5229 Chain = Val.getValue(1); 5230 InFlag = Val.getValue(2); 5231 CopiedRegs[VA.getLocReg()] = Val; 5232 } 5233 5234 switch (VA.getLocInfo()) { 5235 default: 5236 llvm_unreachable("Unknown loc info!"); 5237 case CCValAssign::Full: 5238 break; 5239 case CCValAssign::BCvt: 5240 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val); 5241 break; 5242 case CCValAssign::AExtUpper: 5243 Val = DAG.getNode(ISD::SRL, DL, VA.getLocVT(), Val, 5244 DAG.getConstant(32, DL, VA.getLocVT())); 5245 LLVM_FALLTHROUGH; 5246 case CCValAssign::AExt: 5247 LLVM_FALLTHROUGH; 5248 case CCValAssign::ZExt: 5249 Val = DAG.getZExtOrTrunc(Val, DL, VA.getValVT()); 5250 break; 5251 } 5252 5253 InVals.push_back(Val); 5254 } 5255 5256 return Chain; 5257 } 5258 5259 /// Return true if the calling convention is one that we can guarantee TCO for. 5260 static bool canGuaranteeTCO(CallingConv::ID CC, bool GuaranteeTailCalls) { 5261 return (CC == CallingConv::Fast && GuaranteeTailCalls) || 5262 CC == CallingConv::Tail || CC == CallingConv::SwiftTail; 5263 } 5264 5265 /// Return true if we might ever do TCO for calls with this calling convention. 5266 static bool mayTailCallThisCC(CallingConv::ID CC) { 5267 switch (CC) { 5268 case CallingConv::C: 5269 case CallingConv::AArch64_SVE_VectorCall: 5270 case CallingConv::PreserveMost: 5271 case CallingConv::Swift: 5272 case CallingConv::SwiftTail: 5273 case CallingConv::Tail: 5274 case CallingConv::Fast: 5275 return true; 5276 default: 5277 return false; 5278 } 5279 } 5280 5281 bool AArch64TargetLowering::isEligibleForTailCallOptimization( 5282 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg, 5283 const SmallVectorImpl<ISD::OutputArg> &Outs, 5284 const SmallVectorImpl<SDValue> &OutVals, 5285 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const { 5286 if (!mayTailCallThisCC(CalleeCC)) 5287 return false; 5288 5289 MachineFunction &MF = DAG.getMachineFunction(); 5290 const Function &CallerF = MF.getFunction(); 5291 CallingConv::ID CallerCC = CallerF.getCallingConv(); 5292 5293 // Functions using the C or Fast calling convention that have an SVE signature 5294 // preserve more registers and should assume the SVE_VectorCall CC. 5295 // The check for matching callee-saved regs will determine whether it is 5296 // eligible for TCO. 5297 if ((CallerCC == CallingConv::C || CallerCC == CallingConv::Fast) && 5298 AArch64RegisterInfo::hasSVEArgsOrReturn(&MF)) 5299 CallerCC = CallingConv::AArch64_SVE_VectorCall; 5300 5301 bool CCMatch = CallerCC == CalleeCC; 5302 5303 // When using the Windows calling convention on a non-windows OS, we want 5304 // to back up and restore X18 in such functions; we can't do a tail call 5305 // from those functions. 5306 if (CallerCC == CallingConv::Win64 && !Subtarget->isTargetWindows() && 5307 CalleeCC != CallingConv::Win64) 5308 return false; 5309 5310 // Byval parameters hand the function a pointer directly into the stack area 5311 // we want to reuse during a tail call. Working around this *is* possible (see 5312 // X86) but less efficient and uglier in LowerCall. 5313 for (Function::const_arg_iterator i = CallerF.arg_begin(), 5314 e = CallerF.arg_end(); 5315 i != e; ++i) { 5316 if (i->hasByValAttr()) 5317 return false; 5318 5319 // On Windows, "inreg" attributes signify non-aggregate indirect returns. 5320 // In this case, it is necessary to save/restore X0 in the callee. Tail 5321 // call opt interferes with this. So we disable tail call opt when the 5322 // caller has an argument with "inreg" attribute. 5323 5324 // FIXME: Check whether the callee also has an "inreg" argument. 5325 if (i->hasInRegAttr()) 5326 return false; 5327 } 5328 5329 if (canGuaranteeTCO(CalleeCC, getTargetMachine().Options.GuaranteedTailCallOpt)) 5330 return CCMatch; 5331 5332 // Externally-defined functions with weak linkage should not be 5333 // tail-called on AArch64 when the OS does not support dynamic 5334 // pre-emption of symbols, as the AAELF spec requires normal calls 5335 // to undefined weak functions to be replaced with a NOP or jump to the 5336 // next instruction. The behaviour of branch instructions in this 5337 // situation (as used for tail calls) is implementation-defined, so we 5338 // cannot rely on the linker replacing the tail call with a return. 5339 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 5340 const GlobalValue *GV = G->getGlobal(); 5341 const Triple &TT = getTargetMachine().getTargetTriple(); 5342 if (GV->hasExternalWeakLinkage() && 5343 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO())) 5344 return false; 5345 } 5346 5347 // Now we search for cases where we can use a tail call without changing the 5348 // ABI. Sibcall is used in some places (particularly gcc) to refer to this 5349 // concept. 5350 5351 // I want anyone implementing a new calling convention to think long and hard 5352 // about this assert. 5353 assert((!isVarArg || CalleeCC == CallingConv::C) && 5354 "Unexpected variadic calling convention"); 5355 5356 LLVMContext &C = *DAG.getContext(); 5357 if (isVarArg && !Outs.empty()) { 5358 // At least two cases here: if caller is fastcc then we can't have any 5359 // memory arguments (we'd be expected to clean up the stack afterwards). If 5360 // caller is C then we could potentially use its argument area. 5361 5362 // FIXME: for now we take the most conservative of these in both cases: 5363 // disallow all variadic memory operands. 5364 SmallVector<CCValAssign, 16> ArgLocs; 5365 CCState CCInfo(CalleeCC, isVarArg, MF, ArgLocs, C); 5366 5367 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, true)); 5368 for (const CCValAssign &ArgLoc : ArgLocs) 5369 if (!ArgLoc.isRegLoc()) 5370 return false; 5371 } 5372 5373 // Check that the call results are passed in the same way. 5374 if (!CCState::resultsCompatible(CalleeCC, CallerCC, MF, C, Ins, 5375 CCAssignFnForCall(CalleeCC, isVarArg), 5376 CCAssignFnForCall(CallerCC, isVarArg))) 5377 return false; 5378 // The callee has to preserve all registers the caller needs to preserve. 5379 const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo(); 5380 const uint32_t *CallerPreserved = TRI->getCallPreservedMask(MF, CallerCC); 5381 if (!CCMatch) { 5382 const uint32_t *CalleePreserved = TRI->getCallPreservedMask(MF, CalleeCC); 5383 if (Subtarget->hasCustomCallingConv()) { 5384 TRI->UpdateCustomCallPreservedMask(MF, &CallerPreserved); 5385 TRI->UpdateCustomCallPreservedMask(MF, &CalleePreserved); 5386 } 5387 if (!TRI->regmaskSubsetEqual(CallerPreserved, CalleePreserved)) 5388 return false; 5389 } 5390 5391 // Nothing more to check if the callee is taking no arguments 5392 if (Outs.empty()) 5393 return true; 5394 5395 SmallVector<CCValAssign, 16> ArgLocs; 5396 CCState CCInfo(CalleeCC, isVarArg, MF, ArgLocs, C); 5397 5398 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, isVarArg)); 5399 5400 const AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>(); 5401 5402 // If any of the arguments is passed indirectly, it must be SVE, so the 5403 // 'getBytesInStackArgArea' is not sufficient to determine whether we need to 5404 // allocate space on the stack. That is why we determine this explicitly here 5405 // the call cannot be a tailcall. 5406 if (llvm::any_of(ArgLocs, [](CCValAssign &A) { 5407 assert((A.getLocInfo() != CCValAssign::Indirect || 5408 A.getValVT().isScalableVector()) && 5409 "Expected value to be scalable"); 5410 return A.getLocInfo() == CCValAssign::Indirect; 5411 })) 5412 return false; 5413 5414 // If the stack arguments for this call do not fit into our own save area then 5415 // the call cannot be made tail. 5416 if (CCInfo.getNextStackOffset() > FuncInfo->getBytesInStackArgArea()) 5417 return false; 5418 5419 const MachineRegisterInfo &MRI = MF.getRegInfo(); 5420 if (!parametersInCSRMatch(MRI, CallerPreserved, ArgLocs, OutVals)) 5421 return false; 5422 5423 return true; 5424 } 5425 5426 SDValue AArch64TargetLowering::addTokenForArgument(SDValue Chain, 5427 SelectionDAG &DAG, 5428 MachineFrameInfo &MFI, 5429 int ClobberedFI) const { 5430 SmallVector<SDValue, 8> ArgChains; 5431 int64_t FirstByte = MFI.getObjectOffset(ClobberedFI); 5432 int64_t LastByte = FirstByte + MFI.getObjectSize(ClobberedFI) - 1; 5433 5434 // Include the original chain at the beginning of the list. When this is 5435 // used by target LowerCall hooks, this helps legalize find the 5436 // CALLSEQ_BEGIN node. 5437 ArgChains.push_back(Chain); 5438 5439 // Add a chain value for each stack argument corresponding 5440 for (SDNode::use_iterator U = DAG.getEntryNode().getNode()->use_begin(), 5441 UE = DAG.getEntryNode().getNode()->use_end(); 5442 U != UE; ++U) 5443 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U)) 5444 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr())) 5445 if (FI->getIndex() < 0) { 5446 int64_t InFirstByte = MFI.getObjectOffset(FI->getIndex()); 5447 int64_t InLastByte = InFirstByte; 5448 InLastByte += MFI.getObjectSize(FI->getIndex()) - 1; 5449 5450 if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) || 5451 (FirstByte <= InFirstByte && InFirstByte <= LastByte)) 5452 ArgChains.push_back(SDValue(L, 1)); 5453 } 5454 5455 // Build a tokenfactor for all the chains. 5456 return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains); 5457 } 5458 5459 bool AArch64TargetLowering::DoesCalleeRestoreStack(CallingConv::ID CallCC, 5460 bool TailCallOpt) const { 5461 return (CallCC == CallingConv::Fast && TailCallOpt) || 5462 CallCC == CallingConv::Tail || CallCC == CallingConv::SwiftTail; 5463 } 5464 5465 /// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain, 5466 /// and add input and output parameter nodes. 5467 SDValue 5468 AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI, 5469 SmallVectorImpl<SDValue> &InVals) const { 5470 SelectionDAG &DAG = CLI.DAG; 5471 SDLoc &DL = CLI.DL; 5472 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs; 5473 SmallVector<SDValue, 32> &OutVals = CLI.OutVals; 5474 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins; 5475 SDValue Chain = CLI.Chain; 5476 SDValue Callee = CLI.Callee; 5477 bool &IsTailCall = CLI.IsTailCall; 5478 CallingConv::ID CallConv = CLI.CallConv; 5479 bool IsVarArg = CLI.IsVarArg; 5480 5481 MachineFunction &MF = DAG.getMachineFunction(); 5482 MachineFunction::CallSiteInfo CSInfo; 5483 bool IsThisReturn = false; 5484 5485 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>(); 5486 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt; 5487 bool IsSibCall = false; 5488 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CallConv); 5489 5490 // Check callee args/returns for SVE registers and set calling convention 5491 // accordingly. 5492 if (CallConv == CallingConv::C || CallConv == CallingConv::Fast) { 5493 bool CalleeOutSVE = any_of(Outs, [](ISD::OutputArg &Out){ 5494 return Out.VT.isScalableVector(); 5495 }); 5496 bool CalleeInSVE = any_of(Ins, [](ISD::InputArg &In){ 5497 return In.VT.isScalableVector(); 5498 }); 5499 5500 if (CalleeInSVE || CalleeOutSVE) 5501 CallConv = CallingConv::AArch64_SVE_VectorCall; 5502 } 5503 5504 if (IsTailCall) { 5505 // Check if it's really possible to do a tail call. 5506 IsTailCall = isEligibleForTailCallOptimization( 5507 Callee, CallConv, IsVarArg, Outs, OutVals, Ins, DAG); 5508 if (!IsTailCall && CLI.CB && CLI.CB->isMustTailCall()) 5509 report_fatal_error("failed to perform tail call elimination on a call " 5510 "site marked musttail"); 5511 5512 // A sibling call is one where we're under the usual C ABI and not planning 5513 // to change that but can still do a tail call: 5514 if (!TailCallOpt && IsTailCall && CallConv != CallingConv::Tail && 5515 CallConv != CallingConv::SwiftTail) 5516 IsSibCall = true; 5517 5518 if (IsTailCall) 5519 ++NumTailCalls; 5520 } 5521 5522 // Analyze operands of the call, assigning locations to each operand. 5523 SmallVector<CCValAssign, 16> ArgLocs; 5524 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs, 5525 *DAG.getContext()); 5526 5527 if (IsVarArg) { 5528 // Handle fixed and variable vector arguments differently. 5529 // Variable vector arguments always go into memory. 5530 unsigned NumArgs = Outs.size(); 5531 5532 for (unsigned i = 0; i != NumArgs; ++i) { 5533 MVT ArgVT = Outs[i].VT; 5534 if (!Outs[i].IsFixed && ArgVT.isScalableVector()) 5535 report_fatal_error("Passing SVE types to variadic functions is " 5536 "currently not supported"); 5537 5538 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags; 5539 bool UseVarArgCC = !Outs[i].IsFixed; 5540 // On Windows, the fixed arguments in a vararg call are passed in GPRs 5541 // too, so use the vararg CC to force them to integer registers. 5542 if (IsCalleeWin64) 5543 UseVarArgCC = true; 5544 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, UseVarArgCC); 5545 bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo); 5546 assert(!Res && "Call operand has unhandled type"); 5547 (void)Res; 5548 } 5549 } else { 5550 // At this point, Outs[].VT may already be promoted to i32. To correctly 5551 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and 5552 // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT. 5553 // Since AnalyzeCallOperands uses Ins[].VT for both ValVT and LocVT, here 5554 // we use a special version of AnalyzeCallOperands to pass in ValVT and 5555 // LocVT. 5556 unsigned NumArgs = Outs.size(); 5557 for (unsigned i = 0; i != NumArgs; ++i) { 5558 MVT ValVT = Outs[i].VT; 5559 // Get type of the original argument. 5560 EVT ActualVT = getValueType(DAG.getDataLayout(), 5561 CLI.getArgs()[Outs[i].OrigArgIndex].Ty, 5562 /*AllowUnknown*/ true); 5563 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ValVT; 5564 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags; 5565 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16. 5566 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8) 5567 ValVT = MVT::i8; 5568 else if (ActualMVT == MVT::i16) 5569 ValVT = MVT::i16; 5570 5571 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false); 5572 bool Res = AssignFn(i, ValVT, ValVT, CCValAssign::Full, ArgFlags, CCInfo); 5573 assert(!Res && "Call operand has unhandled type"); 5574 (void)Res; 5575 } 5576 } 5577 5578 // Get a count of how many bytes are to be pushed on the stack. 5579 unsigned NumBytes = CCInfo.getNextStackOffset(); 5580 5581 if (IsSibCall) { 5582 // Since we're not changing the ABI to make this a tail call, the memory 5583 // operands are already available in the caller's incoming argument space. 5584 NumBytes = 0; 5585 } 5586 5587 // FPDiff is the byte offset of the call's argument area from the callee's. 5588 // Stores to callee stack arguments will be placed in FixedStackSlots offset 5589 // by this amount for a tail call. In a sibling call it must be 0 because the 5590 // caller will deallocate the entire stack and the callee still expects its 5591 // arguments to begin at SP+0. Completely unused for non-tail calls. 5592 int FPDiff = 0; 5593 5594 if (IsTailCall && !IsSibCall) { 5595 unsigned NumReusableBytes = FuncInfo->getBytesInStackArgArea(); 5596 5597 // Since callee will pop argument stack as a tail call, we must keep the 5598 // popped size 16-byte aligned. 5599 NumBytes = alignTo(NumBytes, 16); 5600 5601 // FPDiff will be negative if this tail call requires more space than we 5602 // would automatically have in our incoming argument space. Positive if we 5603 // can actually shrink the stack. 5604 FPDiff = NumReusableBytes - NumBytes; 5605 5606 // Update the required reserved area if this is the tail call requiring the 5607 // most argument stack space. 5608 if (FPDiff < 0 && FuncInfo->getTailCallReservedStack() < (unsigned)-FPDiff) 5609 FuncInfo->setTailCallReservedStack(-FPDiff); 5610 5611 // The stack pointer must be 16-byte aligned at all times it's used for a 5612 // memory operation, which in practice means at *all* times and in 5613 // particular across call boundaries. Therefore our own arguments started at 5614 // a 16-byte aligned SP and the delta applied for the tail call should 5615 // satisfy the same constraint. 5616 assert(FPDiff % 16 == 0 && "unaligned stack on tail call"); 5617 } 5618 5619 // Adjust the stack pointer for the new arguments... 5620 // These operations are automatically eliminated by the prolog/epilog pass 5621 if (!IsSibCall) 5622 Chain = DAG.getCALLSEQ_START(Chain, IsTailCall ? 0 : NumBytes, 0, DL); 5623 5624 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, AArch64::SP, 5625 getPointerTy(DAG.getDataLayout())); 5626 5627 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass; 5628 SmallSet<unsigned, 8> RegsUsed; 5629 SmallVector<SDValue, 8> MemOpChains; 5630 auto PtrVT = getPointerTy(DAG.getDataLayout()); 5631 5632 if (IsVarArg && CLI.CB && CLI.CB->isMustTailCall()) { 5633 const auto &Forwards = FuncInfo->getForwardedMustTailRegParms(); 5634 for (const auto &F : Forwards) { 5635 SDValue Val = DAG.getCopyFromReg(Chain, DL, F.VReg, F.VT); 5636 RegsToPass.emplace_back(F.PReg, Val); 5637 } 5638 } 5639 5640 // Walk the register/memloc assignments, inserting copies/loads. 5641 unsigned ExtraArgLocs = 0; 5642 for (unsigned i = 0, e = Outs.size(); i != e; ++i) { 5643 CCValAssign &VA = ArgLocs[i - ExtraArgLocs]; 5644 SDValue Arg = OutVals[i]; 5645 ISD::ArgFlagsTy Flags = Outs[i].Flags; 5646 5647 // Promote the value if needed. 5648 switch (VA.getLocInfo()) { 5649 default: 5650 llvm_unreachable("Unknown loc info!"); 5651 case CCValAssign::Full: 5652 break; 5653 case CCValAssign::SExt: 5654 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg); 5655 break; 5656 case CCValAssign::ZExt: 5657 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg); 5658 break; 5659 case CCValAssign::AExt: 5660 if (Outs[i].ArgVT == MVT::i1) { 5661 // AAPCS requires i1 to be zero-extended to 8-bits by the caller. 5662 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg); 5663 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i8, Arg); 5664 } 5665 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg); 5666 break; 5667 case CCValAssign::AExtUpper: 5668 assert(VA.getValVT() == MVT::i32 && "only expect 32 -> 64 upper bits"); 5669 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg); 5670 Arg = DAG.getNode(ISD::SHL, DL, VA.getLocVT(), Arg, 5671 DAG.getConstant(32, DL, VA.getLocVT())); 5672 break; 5673 case CCValAssign::BCvt: 5674 Arg = DAG.getBitcast(VA.getLocVT(), Arg); 5675 break; 5676 case CCValAssign::Trunc: 5677 Arg = DAG.getZExtOrTrunc(Arg, DL, VA.getLocVT()); 5678 break; 5679 case CCValAssign::FPExt: 5680 Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg); 5681 break; 5682 case CCValAssign::Indirect: 5683 assert(VA.getValVT().isScalableVector() && 5684 "Only scalable vectors can be passed indirectly"); 5685 5686 uint64_t StoreSize = VA.getValVT().getStoreSize().getKnownMinSize(); 5687 uint64_t PartSize = StoreSize; 5688 unsigned NumParts = 1; 5689 if (Outs[i].Flags.isInConsecutiveRegs()) { 5690 assert(!Outs[i].Flags.isInConsecutiveRegsLast()); 5691 while (!Outs[i + NumParts - 1].Flags.isInConsecutiveRegsLast()) 5692 ++NumParts; 5693 StoreSize *= NumParts; 5694 } 5695 5696 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); 5697 Type *Ty = EVT(VA.getValVT()).getTypeForEVT(*DAG.getContext()); 5698 Align Alignment = DAG.getDataLayout().getPrefTypeAlign(Ty); 5699 int FI = MFI.CreateStackObject(StoreSize, Alignment, false); 5700 MFI.setStackID(FI, TargetStackID::ScalableVector); 5701 5702 MachinePointerInfo MPI = 5703 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI); 5704 SDValue Ptr = DAG.getFrameIndex( 5705 FI, DAG.getTargetLoweringInfo().getFrameIndexTy(DAG.getDataLayout())); 5706 SDValue SpillSlot = Ptr; 5707 5708 // Ensure we generate all stores for each tuple part, whilst updating the 5709 // pointer after each store correctly using vscale. 5710 while (NumParts) { 5711 Chain = DAG.getStore(Chain, DL, OutVals[i], Ptr, MPI); 5712 NumParts--; 5713 if (NumParts > 0) { 5714 SDValue BytesIncrement = DAG.getVScale( 5715 DL, Ptr.getValueType(), 5716 APInt(Ptr.getValueSizeInBits().getFixedSize(), PartSize)); 5717 SDNodeFlags Flags; 5718 Flags.setNoUnsignedWrap(true); 5719 5720 MPI = MachinePointerInfo(MPI.getAddrSpace()); 5721 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr, 5722 BytesIncrement, Flags); 5723 ExtraArgLocs++; 5724 i++; 5725 } 5726 } 5727 5728 Arg = SpillSlot; 5729 break; 5730 } 5731 5732 if (VA.isRegLoc()) { 5733 if (i == 0 && Flags.isReturned() && !Flags.isSwiftSelf() && 5734 Outs[0].VT == MVT::i64) { 5735 assert(VA.getLocVT() == MVT::i64 && 5736 "unexpected calling convention register assignment"); 5737 assert(!Ins.empty() && Ins[0].VT == MVT::i64 && 5738 "unexpected use of 'returned'"); 5739 IsThisReturn = true; 5740 } 5741 if (RegsUsed.count(VA.getLocReg())) { 5742 // If this register has already been used then we're trying to pack 5743 // parts of an [N x i32] into an X-register. The extension type will 5744 // take care of putting the two halves in the right place but we have to 5745 // combine them. 5746 SDValue &Bits = 5747 llvm::find_if(RegsToPass, 5748 [=](const std::pair<unsigned, SDValue> &Elt) { 5749 return Elt.first == VA.getLocReg(); 5750 }) 5751 ->second; 5752 Bits = DAG.getNode(ISD::OR, DL, Bits.getValueType(), Bits, Arg); 5753 // Call site info is used for function's parameter entry value 5754 // tracking. For now we track only simple cases when parameter 5755 // is transferred through whole register. 5756 llvm::erase_if(CSInfo, [&VA](MachineFunction::ArgRegPair ArgReg) { 5757 return ArgReg.Reg == VA.getLocReg(); 5758 }); 5759 } else { 5760 RegsToPass.emplace_back(VA.getLocReg(), Arg); 5761 RegsUsed.insert(VA.getLocReg()); 5762 const TargetOptions &Options = DAG.getTarget().Options; 5763 if (Options.EmitCallSiteInfo) 5764 CSInfo.emplace_back(VA.getLocReg(), i); 5765 } 5766 } else { 5767 assert(VA.isMemLoc()); 5768 5769 SDValue DstAddr; 5770 MachinePointerInfo DstInfo; 5771 5772 // FIXME: This works on big-endian for composite byvals, which are the 5773 // common case. It should also work for fundamental types too. 5774 uint32_t BEAlign = 0; 5775 unsigned OpSize; 5776 if (VA.getLocInfo() == CCValAssign::Indirect) 5777 OpSize = VA.getLocVT().getFixedSizeInBits(); 5778 else 5779 OpSize = Flags.isByVal() ? Flags.getByValSize() * 8 5780 : VA.getValVT().getSizeInBits(); 5781 OpSize = (OpSize + 7) / 8; 5782 if (!Subtarget->isLittleEndian() && !Flags.isByVal() && 5783 !Flags.isInConsecutiveRegs()) { 5784 if (OpSize < 8) 5785 BEAlign = 8 - OpSize; 5786 } 5787 unsigned LocMemOffset = VA.getLocMemOffset(); 5788 int32_t Offset = LocMemOffset + BEAlign; 5789 SDValue PtrOff = DAG.getIntPtrConstant(Offset, DL); 5790 PtrOff = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, PtrOff); 5791 5792 if (IsTailCall) { 5793 Offset = Offset + FPDiff; 5794 int FI = MF.getFrameInfo().CreateFixedObject(OpSize, Offset, true); 5795 5796 DstAddr = DAG.getFrameIndex(FI, PtrVT); 5797 DstInfo = 5798 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI); 5799 5800 // Make sure any stack arguments overlapping with where we're storing 5801 // are loaded before this eventual operation. Otherwise they'll be 5802 // clobbered. 5803 Chain = addTokenForArgument(Chain, DAG, MF.getFrameInfo(), FI); 5804 } else { 5805 SDValue PtrOff = DAG.getIntPtrConstant(Offset, DL); 5806 5807 DstAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, PtrOff); 5808 DstInfo = MachinePointerInfo::getStack(DAG.getMachineFunction(), 5809 LocMemOffset); 5810 } 5811 5812 if (Outs[i].Flags.isByVal()) { 5813 SDValue SizeNode = 5814 DAG.getConstant(Outs[i].Flags.getByValSize(), DL, MVT::i64); 5815 SDValue Cpy = DAG.getMemcpy( 5816 Chain, DL, DstAddr, Arg, SizeNode, 5817 Outs[i].Flags.getNonZeroByValAlign(), 5818 /*isVol = */ false, /*AlwaysInline = */ false, 5819 /*isTailCall = */ false, DstInfo, MachinePointerInfo()); 5820 5821 MemOpChains.push_back(Cpy); 5822 } else { 5823 // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already 5824 // promoted to a legal register type i32, we should truncate Arg back to 5825 // i1/i8/i16. 5826 if (VA.getValVT() == MVT::i1 || VA.getValVT() == MVT::i8 || 5827 VA.getValVT() == MVT::i16) 5828 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Arg); 5829 5830 SDValue Store = DAG.getStore(Chain, DL, Arg, DstAddr, DstInfo); 5831 MemOpChains.push_back(Store); 5832 } 5833 } 5834 } 5835 5836 if (!MemOpChains.empty()) 5837 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains); 5838 5839 // Build a sequence of copy-to-reg nodes chained together with token chain 5840 // and flag operands which copy the outgoing args into the appropriate regs. 5841 SDValue InFlag; 5842 for (auto &RegToPass : RegsToPass) { 5843 Chain = DAG.getCopyToReg(Chain, DL, RegToPass.first, 5844 RegToPass.second, InFlag); 5845 InFlag = Chain.getValue(1); 5846 } 5847 5848 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every 5849 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol 5850 // node so that legalize doesn't hack it. 5851 if (auto *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 5852 auto GV = G->getGlobal(); 5853 unsigned OpFlags = 5854 Subtarget->classifyGlobalFunctionReference(GV, getTargetMachine()); 5855 if (OpFlags & AArch64II::MO_GOT) { 5856 Callee = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags); 5857 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, Callee); 5858 } else { 5859 const GlobalValue *GV = G->getGlobal(); 5860 Callee = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 0); 5861 } 5862 } else if (auto *S = dyn_cast<ExternalSymbolSDNode>(Callee)) { 5863 if (getTargetMachine().getCodeModel() == CodeModel::Large && 5864 Subtarget->isTargetMachO()) { 5865 const char *Sym = S->getSymbol(); 5866 Callee = DAG.getTargetExternalSymbol(Sym, PtrVT, AArch64II::MO_GOT); 5867 Callee = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, Callee); 5868 } else { 5869 const char *Sym = S->getSymbol(); 5870 Callee = DAG.getTargetExternalSymbol(Sym, PtrVT, 0); 5871 } 5872 } 5873 5874 // We don't usually want to end the call-sequence here because we would tidy 5875 // the frame up *after* the call, however in the ABI-changing tail-call case 5876 // we've carefully laid out the parameters so that when sp is reset they'll be 5877 // in the correct location. 5878 if (IsTailCall && !IsSibCall) { 5879 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, DL, true), 5880 DAG.getIntPtrConstant(0, DL, true), InFlag, DL); 5881 InFlag = Chain.getValue(1); 5882 } 5883 5884 std::vector<SDValue> Ops; 5885 Ops.push_back(Chain); 5886 Ops.push_back(Callee); 5887 5888 if (IsTailCall) { 5889 // Each tail call may have to adjust the stack by a different amount, so 5890 // this information must travel along with the operation for eventual 5891 // consumption by emitEpilogue. 5892 Ops.push_back(DAG.getTargetConstant(FPDiff, DL, MVT::i32)); 5893 } 5894 5895 // Add argument registers to the end of the list so that they are known live 5896 // into the call. 5897 for (auto &RegToPass : RegsToPass) 5898 Ops.push_back(DAG.getRegister(RegToPass.first, 5899 RegToPass.second.getValueType())); 5900 5901 // Add a register mask operand representing the call-preserved registers. 5902 const uint32_t *Mask; 5903 const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo(); 5904 if (IsThisReturn) { 5905 // For 'this' returns, use the X0-preserving mask if applicable 5906 Mask = TRI->getThisReturnPreservedMask(MF, CallConv); 5907 if (!Mask) { 5908 IsThisReturn = false; 5909 Mask = TRI->getCallPreservedMask(MF, CallConv); 5910 } 5911 } else 5912 Mask = TRI->getCallPreservedMask(MF, CallConv); 5913 5914 if (Subtarget->hasCustomCallingConv()) 5915 TRI->UpdateCustomCallPreservedMask(MF, &Mask); 5916 5917 if (TRI->isAnyArgRegReserved(MF)) 5918 TRI->emitReservedArgRegCallError(MF); 5919 5920 assert(Mask && "Missing call preserved mask for calling convention"); 5921 Ops.push_back(DAG.getRegisterMask(Mask)); 5922 5923 if (InFlag.getNode()) 5924 Ops.push_back(InFlag); 5925 5926 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); 5927 5928 // If we're doing a tall call, use a TC_RETURN here rather than an 5929 // actual call instruction. 5930 if (IsTailCall) { 5931 MF.getFrameInfo().setHasTailCall(); 5932 SDValue Ret = DAG.getNode(AArch64ISD::TC_RETURN, DL, NodeTys, Ops); 5933 DAG.addCallSiteInfo(Ret.getNode(), std::move(CSInfo)); 5934 return Ret; 5935 } 5936 5937 unsigned CallOpc = AArch64ISD::CALL; 5938 // Calls with operand bundle "clang.arc.attachedcall" are special. They should 5939 // be expanded to the call, directly followed by a special marker sequence. 5940 // Use the CALL_RVMARKER to do that. 5941 if (CLI.CB && objcarc::hasAttachedCallOpBundle(CLI.CB)) { 5942 assert(!IsTailCall && 5943 "tail calls cannot be marked with clang.arc.attachedcall"); 5944 CallOpc = AArch64ISD::CALL_RVMARKER; 5945 } 5946 5947 // Returns a chain and a flag for retval copy to use. 5948 Chain = DAG.getNode(CallOpc, DL, NodeTys, Ops); 5949 DAG.addNoMergeSiteInfo(Chain.getNode(), CLI.NoMerge); 5950 InFlag = Chain.getValue(1); 5951 DAG.addCallSiteInfo(Chain.getNode(), std::move(CSInfo)); 5952 5953 uint64_t CalleePopBytes = 5954 DoesCalleeRestoreStack(CallConv, TailCallOpt) ? alignTo(NumBytes, 16) : 0; 5955 5956 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, DL, true), 5957 DAG.getIntPtrConstant(CalleePopBytes, DL, true), 5958 InFlag, DL); 5959 if (!Ins.empty()) 5960 InFlag = Chain.getValue(1); 5961 5962 // Handle result values, copying them out of physregs into vregs that we 5963 // return. 5964 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG, 5965 InVals, IsThisReturn, 5966 IsThisReturn ? OutVals[0] : SDValue()); 5967 } 5968 5969 bool AArch64TargetLowering::CanLowerReturn( 5970 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg, 5971 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const { 5972 CCAssignFn *RetCC = CCAssignFnForReturn(CallConv); 5973 SmallVector<CCValAssign, 16> RVLocs; 5974 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context); 5975 return CCInfo.CheckReturn(Outs, RetCC); 5976 } 5977 5978 SDValue 5979 AArch64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv, 5980 bool isVarArg, 5981 const SmallVectorImpl<ISD::OutputArg> &Outs, 5982 const SmallVectorImpl<SDValue> &OutVals, 5983 const SDLoc &DL, SelectionDAG &DAG) const { 5984 auto &MF = DAG.getMachineFunction(); 5985 auto *FuncInfo = MF.getInfo<AArch64FunctionInfo>(); 5986 5987 CCAssignFn *RetCC = CCAssignFnForReturn(CallConv); 5988 SmallVector<CCValAssign, 16> RVLocs; 5989 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs, 5990 *DAG.getContext()); 5991 CCInfo.AnalyzeReturn(Outs, RetCC); 5992 5993 // Copy the result values into the output registers. 5994 SDValue Flag; 5995 SmallVector<std::pair<unsigned, SDValue>, 4> RetVals; 5996 SmallSet<unsigned, 4> RegsUsed; 5997 for (unsigned i = 0, realRVLocIdx = 0; i != RVLocs.size(); 5998 ++i, ++realRVLocIdx) { 5999 CCValAssign &VA = RVLocs[i]; 6000 assert(VA.isRegLoc() && "Can only return in registers!"); 6001 SDValue Arg = OutVals[realRVLocIdx]; 6002 6003 switch (VA.getLocInfo()) { 6004 default: 6005 llvm_unreachable("Unknown loc info!"); 6006 case CCValAssign::Full: 6007 if (Outs[i].ArgVT == MVT::i1) { 6008 // AAPCS requires i1 to be zero-extended to i8 by the producer of the 6009 // value. This is strictly redundant on Darwin (which uses "zeroext 6010 // i1"), but will be optimised out before ISel. 6011 Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg); 6012 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg); 6013 } 6014 break; 6015 case CCValAssign::BCvt: 6016 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg); 6017 break; 6018 case CCValAssign::AExt: 6019 case CCValAssign::ZExt: 6020 Arg = DAG.getZExtOrTrunc(Arg, DL, VA.getLocVT()); 6021 break; 6022 case CCValAssign::AExtUpper: 6023 assert(VA.getValVT() == MVT::i32 && "only expect 32 -> 64 upper bits"); 6024 Arg = DAG.getZExtOrTrunc(Arg, DL, VA.getLocVT()); 6025 Arg = DAG.getNode(ISD::SHL, DL, VA.getLocVT(), Arg, 6026 DAG.getConstant(32, DL, VA.getLocVT())); 6027 break; 6028 } 6029 6030 if (RegsUsed.count(VA.getLocReg())) { 6031 SDValue &Bits = 6032 llvm::find_if(RetVals, [=](const std::pair<unsigned, SDValue> &Elt) { 6033 return Elt.first == VA.getLocReg(); 6034 })->second; 6035 Bits = DAG.getNode(ISD::OR, DL, Bits.getValueType(), Bits, Arg); 6036 } else { 6037 RetVals.emplace_back(VA.getLocReg(), Arg); 6038 RegsUsed.insert(VA.getLocReg()); 6039 } 6040 } 6041 6042 SmallVector<SDValue, 4> RetOps(1, Chain); 6043 for (auto &RetVal : RetVals) { 6044 Chain = DAG.getCopyToReg(Chain, DL, RetVal.first, RetVal.second, Flag); 6045 Flag = Chain.getValue(1); 6046 RetOps.push_back( 6047 DAG.getRegister(RetVal.first, RetVal.second.getValueType())); 6048 } 6049 6050 // Windows AArch64 ABIs require that for returning structs by value we copy 6051 // the sret argument into X0 for the return. 6052 // We saved the argument into a virtual register in the entry block, 6053 // so now we copy the value out and into X0. 6054 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) { 6055 SDValue Val = DAG.getCopyFromReg(RetOps[0], DL, SRetReg, 6056 getPointerTy(MF.getDataLayout())); 6057 6058 unsigned RetValReg = AArch64::X0; 6059 Chain = DAG.getCopyToReg(Chain, DL, RetValReg, Val, Flag); 6060 Flag = Chain.getValue(1); 6061 6062 RetOps.push_back( 6063 DAG.getRegister(RetValReg, getPointerTy(DAG.getDataLayout()))); 6064 } 6065 6066 const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo(); 6067 const MCPhysReg *I = 6068 TRI->getCalleeSavedRegsViaCopy(&DAG.getMachineFunction()); 6069 if (I) { 6070 for (; *I; ++I) { 6071 if (AArch64::GPR64RegClass.contains(*I)) 6072 RetOps.push_back(DAG.getRegister(*I, MVT::i64)); 6073 else if (AArch64::FPR64RegClass.contains(*I)) 6074 RetOps.push_back(DAG.getRegister(*I, MVT::getFloatingPointVT(64))); 6075 else 6076 llvm_unreachable("Unexpected register class in CSRsViaCopy!"); 6077 } 6078 } 6079 6080 RetOps[0] = Chain; // Update chain. 6081 6082 // Add the flag if we have it. 6083 if (Flag.getNode()) 6084 RetOps.push_back(Flag); 6085 6086 return DAG.getNode(AArch64ISD::RET_FLAG, DL, MVT::Other, RetOps); 6087 } 6088 6089 //===----------------------------------------------------------------------===// 6090 // Other Lowering Code 6091 //===----------------------------------------------------------------------===// 6092 6093 SDValue AArch64TargetLowering::getTargetNode(GlobalAddressSDNode *N, EVT Ty, 6094 SelectionDAG &DAG, 6095 unsigned Flag) const { 6096 return DAG.getTargetGlobalAddress(N->getGlobal(), SDLoc(N), Ty, 6097 N->getOffset(), Flag); 6098 } 6099 6100 SDValue AArch64TargetLowering::getTargetNode(JumpTableSDNode *N, EVT Ty, 6101 SelectionDAG &DAG, 6102 unsigned Flag) const { 6103 return DAG.getTargetJumpTable(N->getIndex(), Ty, Flag); 6104 } 6105 6106 SDValue AArch64TargetLowering::getTargetNode(ConstantPoolSDNode *N, EVT Ty, 6107 SelectionDAG &DAG, 6108 unsigned Flag) const { 6109 return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlign(), 6110 N->getOffset(), Flag); 6111 } 6112 6113 SDValue AArch64TargetLowering::getTargetNode(BlockAddressSDNode* N, EVT Ty, 6114 SelectionDAG &DAG, 6115 unsigned Flag) const { 6116 return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, 0, Flag); 6117 } 6118 6119 // (loadGOT sym) 6120 template <class NodeTy> 6121 SDValue AArch64TargetLowering::getGOT(NodeTy *N, SelectionDAG &DAG, 6122 unsigned Flags) const { 6123 LLVM_DEBUG(dbgs() << "AArch64TargetLowering::getGOT\n"); 6124 SDLoc DL(N); 6125 EVT Ty = getPointerTy(DAG.getDataLayout()); 6126 SDValue GotAddr = getTargetNode(N, Ty, DAG, AArch64II::MO_GOT | Flags); 6127 // FIXME: Once remat is capable of dealing with instructions with register 6128 // operands, expand this into two nodes instead of using a wrapper node. 6129 return DAG.getNode(AArch64ISD::LOADgot, DL, Ty, GotAddr); 6130 } 6131 6132 // (wrapper %highest(sym), %higher(sym), %hi(sym), %lo(sym)) 6133 template <class NodeTy> 6134 SDValue AArch64TargetLowering::getAddrLarge(NodeTy *N, SelectionDAG &DAG, 6135 unsigned Flags) const { 6136 LLVM_DEBUG(dbgs() << "AArch64TargetLowering::getAddrLarge\n"); 6137 SDLoc DL(N); 6138 EVT Ty = getPointerTy(DAG.getDataLayout()); 6139 const unsigned char MO_NC = AArch64II::MO_NC; 6140 return DAG.getNode( 6141 AArch64ISD::WrapperLarge, DL, Ty, 6142 getTargetNode(N, Ty, DAG, AArch64II::MO_G3 | Flags), 6143 getTargetNode(N, Ty, DAG, AArch64II::MO_G2 | MO_NC | Flags), 6144 getTargetNode(N, Ty, DAG, AArch64II::MO_G1 | MO_NC | Flags), 6145 getTargetNode(N, Ty, DAG, AArch64II::MO_G0 | MO_NC | Flags)); 6146 } 6147 6148 // (addlow (adrp %hi(sym)) %lo(sym)) 6149 template <class NodeTy> 6150 SDValue AArch64TargetLowering::getAddr(NodeTy *N, SelectionDAG &DAG, 6151 unsigned Flags) const { 6152 LLVM_DEBUG(dbgs() << "AArch64TargetLowering::getAddr\n"); 6153 SDLoc DL(N); 6154 EVT Ty = getPointerTy(DAG.getDataLayout()); 6155 SDValue Hi = getTargetNode(N, Ty, DAG, AArch64II::MO_PAGE | Flags); 6156 SDValue Lo = getTargetNode(N, Ty, DAG, 6157 AArch64II::MO_PAGEOFF | AArch64II::MO_NC | Flags); 6158 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, Ty, Hi); 6159 return DAG.getNode(AArch64ISD::ADDlow, DL, Ty, ADRP, Lo); 6160 } 6161 6162 // (adr sym) 6163 template <class NodeTy> 6164 SDValue AArch64TargetLowering::getAddrTiny(NodeTy *N, SelectionDAG &DAG, 6165 unsigned Flags) const { 6166 LLVM_DEBUG(dbgs() << "AArch64TargetLowering::getAddrTiny\n"); 6167 SDLoc DL(N); 6168 EVT Ty = getPointerTy(DAG.getDataLayout()); 6169 SDValue Sym = getTargetNode(N, Ty, DAG, Flags); 6170 return DAG.getNode(AArch64ISD::ADR, DL, Ty, Sym); 6171 } 6172 6173 SDValue AArch64TargetLowering::LowerGlobalAddress(SDValue Op, 6174 SelectionDAG &DAG) const { 6175 GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op); 6176 const GlobalValue *GV = GN->getGlobal(); 6177 unsigned OpFlags = Subtarget->ClassifyGlobalReference(GV, getTargetMachine()); 6178 6179 if (OpFlags != AArch64II::MO_NO_FLAG) 6180 assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 && 6181 "unexpected offset in global node"); 6182 6183 // This also catches the large code model case for Darwin, and tiny code 6184 // model with got relocations. 6185 if ((OpFlags & AArch64II::MO_GOT) != 0) { 6186 return getGOT(GN, DAG, OpFlags); 6187 } 6188 6189 SDValue Result; 6190 if (getTargetMachine().getCodeModel() == CodeModel::Large) { 6191 Result = getAddrLarge(GN, DAG, OpFlags); 6192 } else if (getTargetMachine().getCodeModel() == CodeModel::Tiny) { 6193 Result = getAddrTiny(GN, DAG, OpFlags); 6194 } else { 6195 Result = getAddr(GN, DAG, OpFlags); 6196 } 6197 EVT PtrVT = getPointerTy(DAG.getDataLayout()); 6198 SDLoc DL(GN); 6199 if (OpFlags & (AArch64II::MO_DLLIMPORT | AArch64II::MO_COFFSTUB)) 6200 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result, 6201 MachinePointerInfo::getGOT(DAG.getMachineFunction())); 6202 return Result; 6203 } 6204 6205 /// Convert a TLS address reference into the correct sequence of loads 6206 /// and calls to compute the variable's address (for Darwin, currently) and 6207 /// return an SDValue containing the final node. 6208 6209 /// Darwin only has one TLS scheme which must be capable of dealing with the 6210 /// fully general situation, in the worst case. This means: 6211 /// + "extern __thread" declaration. 6212 /// + Defined in a possibly unknown dynamic library. 6213 /// 6214 /// The general system is that each __thread variable has a [3 x i64] descriptor 6215 /// which contains information used by the runtime to calculate the address. The 6216 /// only part of this the compiler needs to know about is the first xword, which 6217 /// contains a function pointer that must be called with the address of the 6218 /// entire descriptor in "x0". 6219 /// 6220 /// Since this descriptor may be in a different unit, in general even the 6221 /// descriptor must be accessed via an indirect load. The "ideal" code sequence 6222 /// is: 6223 /// adrp x0, _var@TLVPPAGE 6224 /// ldr x0, [x0, _var@TLVPPAGEOFF] ; x0 now contains address of descriptor 6225 /// ldr x1, [x0] ; x1 contains 1st entry of descriptor, 6226 /// ; the function pointer 6227 /// blr x1 ; Uses descriptor address in x0 6228 /// ; Address of _var is now in x0. 6229 /// 6230 /// If the address of _var's descriptor *is* known to the linker, then it can 6231 /// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for 6232 /// a slight efficiency gain. 6233 SDValue 6234 AArch64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op, 6235 SelectionDAG &DAG) const { 6236 assert(Subtarget->isTargetDarwin() && 6237 "This function expects a Darwin target"); 6238 6239 SDLoc DL(Op); 6240 MVT PtrVT = getPointerTy(DAG.getDataLayout()); 6241 MVT PtrMemVT = getPointerMemTy(DAG.getDataLayout()); 6242 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); 6243 6244 SDValue TLVPAddr = 6245 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS); 6246 SDValue DescAddr = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TLVPAddr); 6247 6248 // The first entry in the descriptor is a function pointer that we must call 6249 // to obtain the address of the variable. 6250 SDValue Chain = DAG.getEntryNode(); 6251 SDValue FuncTLVGet = DAG.getLoad( 6252 PtrMemVT, DL, Chain, DescAddr, 6253 MachinePointerInfo::getGOT(DAG.getMachineFunction()), 6254 Align(PtrMemVT.getSizeInBits() / 8), 6255 MachineMemOperand::MOInvariant | MachineMemOperand::MODereferenceable); 6256 Chain = FuncTLVGet.getValue(1); 6257 6258 // Extend loaded pointer if necessary (i.e. if ILP32) to DAG pointer. 6259 FuncTLVGet = DAG.getZExtOrTrunc(FuncTLVGet, DL, PtrVT); 6260 6261 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); 6262 MFI.setAdjustsStack(true); 6263 6264 // TLS calls preserve all registers except those that absolutely must be 6265 // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be 6266 // silly). 6267 const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo(); 6268 const uint32_t *Mask = TRI->getTLSCallPreservedMask(); 6269 if (Subtarget->hasCustomCallingConv()) 6270 TRI->UpdateCustomCallPreservedMask(DAG.getMachineFunction(), &Mask); 6271 6272 // Finally, we can make the call. This is just a degenerate version of a 6273 // normal AArch64 call node: x0 takes the address of the descriptor, and 6274 // returns the address of the variable in this thread. 6275 Chain = DAG.getCopyToReg(Chain, DL, AArch64::X0, DescAddr, SDValue()); 6276 Chain = 6277 DAG.getNode(AArch64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue), 6278 Chain, FuncTLVGet, DAG.getRegister(AArch64::X0, MVT::i64), 6279 DAG.getRegisterMask(Mask), Chain.getValue(1)); 6280 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Chain.getValue(1)); 6281 } 6282 6283 /// Convert a thread-local variable reference into a sequence of instructions to 6284 /// compute the variable's address for the local exec TLS model of ELF targets. 6285 /// The sequence depends on the maximum TLS area size. 6286 SDValue AArch64TargetLowering::LowerELFTLSLocalExec(const GlobalValue *GV, 6287 SDValue ThreadBase, 6288 const SDLoc &DL, 6289 SelectionDAG &DAG) const { 6290 EVT PtrVT = getPointerTy(DAG.getDataLayout()); 6291 SDValue TPOff, Addr; 6292 6293 switch (DAG.getTarget().Options.TLSSize) { 6294 default: 6295 llvm_unreachable("Unexpected TLS size"); 6296 6297 case 12: { 6298 // mrs x0, TPIDR_EL0 6299 // add x0, x0, :tprel_lo12:a 6300 SDValue Var = DAG.getTargetGlobalAddress( 6301 GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_PAGEOFF); 6302 return SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, ThreadBase, 6303 Var, 6304 DAG.getTargetConstant(0, DL, MVT::i32)), 6305 0); 6306 } 6307 6308 case 24: { 6309 // mrs x0, TPIDR_EL0 6310 // add x0, x0, :tprel_hi12:a 6311 // add x0, x0, :tprel_lo12_nc:a 6312 SDValue HiVar = DAG.getTargetGlobalAddress( 6313 GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_HI12); 6314 SDValue LoVar = DAG.getTargetGlobalAddress( 6315 GV, DL, PtrVT, 0, 6316 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC); 6317 Addr = SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, ThreadBase, 6318 HiVar, 6319 DAG.getTargetConstant(0, DL, MVT::i32)), 6320 0); 6321 return SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, Addr, 6322 LoVar, 6323 DAG.getTargetConstant(0, DL, MVT::i32)), 6324 0); 6325 } 6326 6327 case 32: { 6328 // mrs x1, TPIDR_EL0 6329 // movz x0, #:tprel_g1:a 6330 // movk x0, #:tprel_g0_nc:a 6331 // add x0, x1, x0 6332 SDValue HiVar = DAG.getTargetGlobalAddress( 6333 GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_G1); 6334 SDValue LoVar = DAG.getTargetGlobalAddress( 6335 GV, DL, PtrVT, 0, 6336 AArch64II::MO_TLS | AArch64II::MO_G0 | AArch64II::MO_NC); 6337 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVZXi, DL, PtrVT, HiVar, 6338 DAG.getTargetConstant(16, DL, MVT::i32)), 6339 0); 6340 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVKXi, DL, PtrVT, TPOff, LoVar, 6341 DAG.getTargetConstant(0, DL, MVT::i32)), 6342 0); 6343 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff); 6344 } 6345 6346 case 48: { 6347 // mrs x1, TPIDR_EL0 6348 // movz x0, #:tprel_g2:a 6349 // movk x0, #:tprel_g1_nc:a 6350 // movk x0, #:tprel_g0_nc:a 6351 // add x0, x1, x0 6352 SDValue HiVar = DAG.getTargetGlobalAddress( 6353 GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_G2); 6354 SDValue MiVar = DAG.getTargetGlobalAddress( 6355 GV, DL, PtrVT, 0, 6356 AArch64II::MO_TLS | AArch64II::MO_G1 | AArch64II::MO_NC); 6357 SDValue LoVar = DAG.getTargetGlobalAddress( 6358 GV, DL, PtrVT, 0, 6359 AArch64II::MO_TLS | AArch64II::MO_G0 | AArch64II::MO_NC); 6360 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVZXi, DL, PtrVT, HiVar, 6361 DAG.getTargetConstant(32, DL, MVT::i32)), 6362 0); 6363 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVKXi, DL, PtrVT, TPOff, MiVar, 6364 DAG.getTargetConstant(16, DL, MVT::i32)), 6365 0); 6366 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVKXi, DL, PtrVT, TPOff, LoVar, 6367 DAG.getTargetConstant(0, DL, MVT::i32)), 6368 0); 6369 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff); 6370 } 6371 } 6372 } 6373 6374 /// When accessing thread-local variables under either the general-dynamic or 6375 /// local-dynamic system, we make a "TLS-descriptor" call. The variable will 6376 /// have a descriptor, accessible via a PC-relative ADRP, and whose first entry 6377 /// is a function pointer to carry out the resolution. 6378 /// 6379 /// The sequence is: 6380 /// adrp x0, :tlsdesc:var 6381 /// ldr x1, [x0, #:tlsdesc_lo12:var] 6382 /// add x0, x0, #:tlsdesc_lo12:var 6383 /// .tlsdesccall var 6384 /// blr x1 6385 /// (TPIDR_EL0 offset now in x0) 6386 /// 6387 /// The above sequence must be produced unscheduled, to enable the linker to 6388 /// optimize/relax this sequence. 6389 /// Therefore, a pseudo-instruction (TLSDESC_CALLSEQ) is used to represent the 6390 /// above sequence, and expanded really late in the compilation flow, to ensure 6391 /// the sequence is produced as per above. 6392 SDValue AArch64TargetLowering::LowerELFTLSDescCallSeq(SDValue SymAddr, 6393 const SDLoc &DL, 6394 SelectionDAG &DAG) const { 6395 EVT PtrVT = getPointerTy(DAG.getDataLayout()); 6396 6397 SDValue Chain = DAG.getEntryNode(); 6398 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); 6399 6400 Chain = 6401 DAG.getNode(AArch64ISD::TLSDESC_CALLSEQ, DL, NodeTys, {Chain, SymAddr}); 6402 SDValue Glue = Chain.getValue(1); 6403 6404 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Glue); 6405 } 6406 6407 SDValue 6408 AArch64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op, 6409 SelectionDAG &DAG) const { 6410 assert(Subtarget->isTargetELF() && "This function expects an ELF target"); 6411 6412 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); 6413 6414 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal()); 6415 6416 if (!EnableAArch64ELFLocalDynamicTLSGeneration) { 6417 if (Model == TLSModel::LocalDynamic) 6418 Model = TLSModel::GeneralDynamic; 6419 } 6420 6421 if (getTargetMachine().getCodeModel() == CodeModel::Large && 6422 Model != TLSModel::LocalExec) 6423 report_fatal_error("ELF TLS only supported in small memory model or " 6424 "in local exec TLS model"); 6425 // Different choices can be made for the maximum size of the TLS area for a 6426 // module. For the small address model, the default TLS size is 16MiB and the 6427 // maximum TLS size is 4GiB. 6428 // FIXME: add tiny and large code model support for TLS access models other 6429 // than local exec. We currently generate the same code as small for tiny, 6430 // which may be larger than needed. 6431 6432 SDValue TPOff; 6433 EVT PtrVT = getPointerTy(DAG.getDataLayout()); 6434 SDLoc DL(Op); 6435 const GlobalValue *GV = GA->getGlobal(); 6436 6437 SDValue ThreadBase = DAG.getNode(AArch64ISD::THREAD_POINTER, DL, PtrVT); 6438 6439 if (Model == TLSModel::LocalExec) { 6440 return LowerELFTLSLocalExec(GV, ThreadBase, DL, DAG); 6441 } else if (Model == TLSModel::InitialExec) { 6442 TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS); 6443 TPOff = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TPOff); 6444 } else if (Model == TLSModel::LocalDynamic) { 6445 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS 6446 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate 6447 // the beginning of the module's TLS region, followed by a DTPREL offset 6448 // calculation. 6449 6450 // These accesses will need deduplicating if there's more than one. 6451 AArch64FunctionInfo *MFI = 6452 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>(); 6453 MFI->incNumLocalDynamicTLSAccesses(); 6454 6455 // The call needs a relocation too for linker relaxation. It doesn't make 6456 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of 6457 // the address. 6458 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT, 6459 AArch64II::MO_TLS); 6460 6461 // Now we can calculate the offset from TPIDR_EL0 to this module's 6462 // thread-local area. 6463 TPOff = LowerELFTLSDescCallSeq(SymAddr, DL, DAG); 6464 6465 // Now use :dtprel_whatever: operations to calculate this variable's offset 6466 // in its thread-storage area. 6467 SDValue HiVar = DAG.getTargetGlobalAddress( 6468 GV, DL, MVT::i64, 0, AArch64II::MO_TLS | AArch64II::MO_HI12); 6469 SDValue LoVar = DAG.getTargetGlobalAddress( 6470 GV, DL, MVT::i64, 0, 6471 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC); 6472 6473 TPOff = SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPOff, HiVar, 6474 DAG.getTargetConstant(0, DL, MVT::i32)), 6475 0); 6476 TPOff = SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPOff, LoVar, 6477 DAG.getTargetConstant(0, DL, MVT::i32)), 6478 0); 6479 } else if (Model == TLSModel::GeneralDynamic) { 6480 // The call needs a relocation too for linker relaxation. It doesn't make 6481 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of 6482 // the address. 6483 SDValue SymAddr = 6484 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS); 6485 6486 // Finally we can make a call to calculate the offset from tpidr_el0. 6487 TPOff = LowerELFTLSDescCallSeq(SymAddr, DL, DAG); 6488 } else 6489 llvm_unreachable("Unsupported ELF TLS access model"); 6490 6491 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff); 6492 } 6493 6494 SDValue 6495 AArch64TargetLowering::LowerWindowsGlobalTLSAddress(SDValue Op, 6496 SelectionDAG &DAG) const { 6497 assert(Subtarget->isTargetWindows() && "Windows specific TLS lowering"); 6498 6499 SDValue Chain = DAG.getEntryNode(); 6500 EVT PtrVT = getPointerTy(DAG.getDataLayout()); 6501 SDLoc DL(Op); 6502 6503 SDValue TEB = DAG.getRegister(AArch64::X18, MVT::i64); 6504 6505 // Load the ThreadLocalStoragePointer from the TEB 6506 // A pointer to the TLS array is located at offset 0x58 from the TEB. 6507 SDValue TLSArray = 6508 DAG.getNode(ISD::ADD, DL, PtrVT, TEB, DAG.getIntPtrConstant(0x58, DL)); 6509 TLSArray = DAG.getLoad(PtrVT, DL, Chain, TLSArray, MachinePointerInfo()); 6510 Chain = TLSArray.getValue(1); 6511 6512 // Load the TLS index from the C runtime; 6513 // This does the same as getAddr(), but without having a GlobalAddressSDNode. 6514 // This also does the same as LOADgot, but using a generic i32 load, 6515 // while LOADgot only loads i64. 6516 SDValue TLSIndexHi = 6517 DAG.getTargetExternalSymbol("_tls_index", PtrVT, AArch64II::MO_PAGE); 6518 SDValue TLSIndexLo = DAG.getTargetExternalSymbol( 6519 "_tls_index", PtrVT, AArch64II::MO_PAGEOFF | AArch64II::MO_NC); 6520 SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, TLSIndexHi); 6521 SDValue TLSIndex = 6522 DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, TLSIndexLo); 6523 TLSIndex = DAG.getLoad(MVT::i32, DL, Chain, TLSIndex, MachinePointerInfo()); 6524 Chain = TLSIndex.getValue(1); 6525 6526 // The pointer to the thread's TLS data area is at the TLS Index scaled by 8 6527 // offset into the TLSArray. 6528 TLSIndex = DAG.getNode(ISD::ZERO_EXTEND, DL, PtrVT, TLSIndex); 6529 SDValue Slot = DAG.getNode(ISD::SHL, DL, PtrVT, TLSIndex, 6530 DAG.getConstant(3, DL, PtrVT)); 6531 SDValue TLS = DAG.getLoad(PtrVT, DL, Chain, 6532 DAG.getNode(ISD::ADD, DL, PtrVT, TLSArray, Slot), 6533 MachinePointerInfo()); 6534 Chain = TLS.getValue(1); 6535 6536 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); 6537 const GlobalValue *GV = GA->getGlobal(); 6538 SDValue TGAHi = DAG.getTargetGlobalAddress( 6539 GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_HI12); 6540 SDValue TGALo = DAG.getTargetGlobalAddress( 6541 GV, DL, PtrVT, 0, 6542 AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC); 6543 6544 // Add the offset from the start of the .tls section (section base). 6545 SDValue Addr = 6546 SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TLS, TGAHi, 6547 DAG.getTargetConstant(0, DL, MVT::i32)), 6548 0); 6549 Addr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, Addr, TGALo); 6550 return Addr; 6551 } 6552 6553 SDValue AArch64TargetLowering::LowerGlobalTLSAddress(SDValue Op, 6554 SelectionDAG &DAG) const { 6555 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); 6556 if (DAG.getTarget().useEmulatedTLS()) 6557 return LowerToTLSEmulatedModel(GA, DAG); 6558 6559 if (Subtarget->isTargetDarwin()) 6560 return LowerDarwinGlobalTLSAddress(Op, DAG); 6561 if (Subtarget->isTargetELF()) 6562 return LowerELFGlobalTLSAddress(Op, DAG); 6563 if (Subtarget->isTargetWindows()) 6564 return LowerWindowsGlobalTLSAddress(Op, DAG); 6565 6566 llvm_unreachable("Unexpected platform trying to use TLS"); 6567 } 6568 6569 // Looks through \param Val to determine the bit that can be used to 6570 // check the sign of the value. It returns the unextended value and 6571 // the sign bit position. 6572 std::pair<SDValue, uint64_t> lookThroughSignExtension(SDValue Val) { 6573 if (Val.getOpcode() == ISD::SIGN_EXTEND_INREG) 6574 return {Val.getOperand(0), 6575 cast<VTSDNode>(Val.getOperand(1))->getVT().getFixedSizeInBits() - 6576 1}; 6577 6578 if (Val.getOpcode() == ISD::SIGN_EXTEND) 6579 return {Val.getOperand(0), 6580 Val.getOperand(0)->getValueType(0).getFixedSizeInBits() - 1}; 6581 6582 return {Val, Val.getValueSizeInBits() - 1}; 6583 } 6584 6585 SDValue AArch64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const { 6586 SDValue Chain = Op.getOperand(0); 6587 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get(); 6588 SDValue LHS = Op.getOperand(2); 6589 SDValue RHS = Op.getOperand(3); 6590 SDValue Dest = Op.getOperand(4); 6591 SDLoc dl(Op); 6592 6593 MachineFunction &MF = DAG.getMachineFunction(); 6594 // Speculation tracking/SLH assumes that optimized TB(N)Z/CB(N)Z instructions 6595 // will not be produced, as they are conditional branch instructions that do 6596 // not set flags. 6597 bool ProduceNonFlagSettingCondBr = 6598 !MF.getFunction().hasFnAttribute(Attribute::SpeculativeLoadHardening); 6599 6600 // Handle f128 first, since lowering it will result in comparing the return 6601 // value of a libcall against zero, which is just what the rest of LowerBR_CC 6602 // is expecting to deal with. 6603 if (LHS.getValueType() == MVT::f128) { 6604 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl, LHS, RHS); 6605 6606 // If softenSetCCOperands returned a scalar, we need to compare the result 6607 // against zero to select between true and false values. 6608 if (!RHS.getNode()) { 6609 RHS = DAG.getConstant(0, dl, LHS.getValueType()); 6610 CC = ISD::SETNE; 6611 } 6612 } 6613 6614 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch 6615 // instruction. 6616 if (ISD::isOverflowIntrOpRes(LHS) && isOneConstant(RHS) && 6617 (CC == ISD::SETEQ || CC == ISD::SETNE)) { 6618 // Only lower legal XALUO ops. 6619 if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0))) 6620 return SDValue(); 6621 6622 // The actual operation with overflow check. 6623 AArch64CC::CondCode OFCC; 6624 SDValue Value, Overflow; 6625 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, LHS.getValue(0), DAG); 6626 6627 if (CC == ISD::SETNE) 6628 OFCC = getInvertedCondCode(OFCC); 6629 SDValue CCVal = DAG.getConstant(OFCC, dl, MVT::i32); 6630 6631 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal, 6632 Overflow); 6633 } 6634 6635 if (LHS.getValueType().isInteger()) { 6636 assert((LHS.getValueType() == RHS.getValueType()) && 6637 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64)); 6638 6639 // If the RHS of the comparison is zero, we can potentially fold this 6640 // to a specialized branch. 6641 const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS); 6642 if (RHSC && RHSC->getZExtValue() == 0 && ProduceNonFlagSettingCondBr) { 6643 if (CC == ISD::SETEQ) { 6644 // See if we can use a TBZ to fold in an AND as well. 6645 // TBZ has a smaller branch displacement than CBZ. If the offset is 6646 // out of bounds, a late MI-layer pass rewrites branches. 6647 // 403.gcc is an example that hits this case. 6648 if (LHS.getOpcode() == ISD::AND && 6649 isa<ConstantSDNode>(LHS.getOperand(1)) && 6650 isPowerOf2_64(LHS.getConstantOperandVal(1))) { 6651 SDValue Test = LHS.getOperand(0); 6652 uint64_t Mask = LHS.getConstantOperandVal(1); 6653 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, Test, 6654 DAG.getConstant(Log2_64(Mask), dl, MVT::i64), 6655 Dest); 6656 } 6657 6658 return DAG.getNode(AArch64ISD::CBZ, dl, MVT::Other, Chain, LHS, Dest); 6659 } else if (CC == ISD::SETNE) { 6660 // See if we can use a TBZ to fold in an AND as well. 6661 // TBZ has a smaller branch displacement than CBZ. If the offset is 6662 // out of bounds, a late MI-layer pass rewrites branches. 6663 // 403.gcc is an example that hits this case. 6664 if (LHS.getOpcode() == ISD::AND && 6665 isa<ConstantSDNode>(LHS.getOperand(1)) && 6666 isPowerOf2_64(LHS.getConstantOperandVal(1))) { 6667 SDValue Test = LHS.getOperand(0); 6668 uint64_t Mask = LHS.getConstantOperandVal(1); 6669 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, Test, 6670 DAG.getConstant(Log2_64(Mask), dl, MVT::i64), 6671 Dest); 6672 } 6673 6674 return DAG.getNode(AArch64ISD::CBNZ, dl, MVT::Other, Chain, LHS, Dest); 6675 } else if (CC == ISD::SETLT && LHS.getOpcode() != ISD::AND) { 6676 // Don't combine AND since emitComparison converts the AND to an ANDS 6677 // (a.k.a. TST) and the test in the test bit and branch instruction 6678 // becomes redundant. This would also increase register pressure. 6679 uint64_t SignBitPos; 6680 std::tie(LHS, SignBitPos) = lookThroughSignExtension(LHS); 6681 return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, LHS, 6682 DAG.getConstant(SignBitPos, dl, MVT::i64), Dest); 6683 } 6684 } 6685 if (RHSC && RHSC->getSExtValue() == -1 && CC == ISD::SETGT && 6686 LHS.getOpcode() != ISD::AND && ProduceNonFlagSettingCondBr) { 6687 // Don't combine AND since emitComparison converts the AND to an ANDS 6688 // (a.k.a. TST) and the test in the test bit and branch instruction 6689 // becomes redundant. This would also increase register pressure. 6690 uint64_t SignBitPos; 6691 std::tie(LHS, SignBitPos) = lookThroughSignExtension(LHS); 6692 return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, LHS, 6693 DAG.getConstant(SignBitPos, dl, MVT::i64), Dest); 6694 } 6695 6696 SDValue CCVal; 6697 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl); 6698 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal, 6699 Cmp); 6700 } 6701 6702 assert(LHS.getValueType() == MVT::f16 || LHS.getValueType() == MVT::bf16 || 6703 LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64); 6704 6705 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally 6706 // clean. Some of them require two branches to implement. 6707 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG); 6708 AArch64CC::CondCode CC1, CC2; 6709 changeFPCCToAArch64CC(CC, CC1, CC2); 6710 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32); 6711 SDValue BR1 = 6712 DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CC1Val, Cmp); 6713 if (CC2 != AArch64CC::AL) { 6714 SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32); 6715 return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, BR1, Dest, CC2Val, 6716 Cmp); 6717 } 6718 6719 return BR1; 6720 } 6721 6722 SDValue AArch64TargetLowering::LowerFCOPYSIGN(SDValue Op, 6723 SelectionDAG &DAG) const { 6724 EVT VT = Op.getValueType(); 6725 SDLoc DL(Op); 6726 6727 SDValue In1 = Op.getOperand(0); 6728 SDValue In2 = Op.getOperand(1); 6729 EVT SrcVT = In2.getValueType(); 6730 6731 if (SrcVT.bitsLT(VT)) 6732 In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2); 6733 else if (SrcVT.bitsGT(VT)) 6734 In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2, DAG.getIntPtrConstant(0, DL)); 6735 6736 EVT VecVT; 6737 uint64_t EltMask; 6738 SDValue VecVal1, VecVal2; 6739 6740 auto setVecVal = [&] (int Idx) { 6741 if (!VT.isVector()) { 6742 VecVal1 = DAG.getTargetInsertSubreg(Idx, DL, VecVT, 6743 DAG.getUNDEF(VecVT), In1); 6744 VecVal2 = DAG.getTargetInsertSubreg(Idx, DL, VecVT, 6745 DAG.getUNDEF(VecVT), In2); 6746 } else { 6747 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1); 6748 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2); 6749 } 6750 }; 6751 6752 if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) { 6753 VecVT = (VT == MVT::v2f32 ? MVT::v2i32 : MVT::v4i32); 6754 EltMask = 0x80000000ULL; 6755 setVecVal(AArch64::ssub); 6756 } else if (VT == MVT::f64 || VT == MVT::v2f64) { 6757 VecVT = MVT::v2i64; 6758 6759 // We want to materialize a mask with the high bit set, but the AdvSIMD 6760 // immediate moves cannot materialize that in a single instruction for 6761 // 64-bit elements. Instead, materialize zero and then negate it. 6762 EltMask = 0; 6763 6764 setVecVal(AArch64::dsub); 6765 } else if (VT == MVT::f16 || VT == MVT::v4f16 || VT == MVT::v8f16) { 6766 VecVT = (VT == MVT::v4f16 ? MVT::v4i16 : MVT::v8i16); 6767 EltMask = 0x8000ULL; 6768 setVecVal(AArch64::hsub); 6769 } else { 6770 llvm_unreachable("Invalid type for copysign!"); 6771 } 6772 6773 SDValue BuildVec = DAG.getConstant(EltMask, DL, VecVT); 6774 6775 // If we couldn't materialize the mask above, then the mask vector will be 6776 // the zero vector, and we need to negate it here. 6777 if (VT == MVT::f64 || VT == MVT::v2f64) { 6778 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, BuildVec); 6779 BuildVec = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, BuildVec); 6780 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, BuildVec); 6781 } 6782 6783 SDValue Sel = 6784 DAG.getNode(AArch64ISD::BIT, DL, VecVT, VecVal1, VecVal2, BuildVec); 6785 6786 if (VT == MVT::f16) 6787 return DAG.getTargetExtractSubreg(AArch64::hsub, DL, VT, Sel); 6788 if (VT == MVT::f32) 6789 return DAG.getTargetExtractSubreg(AArch64::ssub, DL, VT, Sel); 6790 else if (VT == MVT::f64) 6791 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, VT, Sel); 6792 else 6793 return DAG.getNode(ISD::BITCAST, DL, VT, Sel); 6794 } 6795 6796 SDValue AArch64TargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const { 6797 if (DAG.getMachineFunction().getFunction().hasFnAttribute( 6798 Attribute::NoImplicitFloat)) 6799 return SDValue(); 6800 6801 if (!Subtarget->hasNEON()) 6802 return SDValue(); 6803 6804 // While there is no integer popcount instruction, it can 6805 // be more efficiently lowered to the following sequence that uses 6806 // AdvSIMD registers/instructions as long as the copies to/from 6807 // the AdvSIMD registers are cheap. 6808 // FMOV D0, X0 // copy 64-bit int to vector, high bits zero'd 6809 // CNT V0.8B, V0.8B // 8xbyte pop-counts 6810 // ADDV B0, V0.8B // sum 8xbyte pop-counts 6811 // UMOV X0, V0.B[0] // copy byte result back to integer reg 6812 SDValue Val = Op.getOperand(0); 6813 SDLoc DL(Op); 6814 EVT VT = Op.getValueType(); 6815 6816 if (VT == MVT::i32 || VT == MVT::i64) { 6817 if (VT == MVT::i32) 6818 Val = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Val); 6819 Val = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val); 6820 6821 SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, Val); 6822 SDValue UaddLV = DAG.getNode( 6823 ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32, 6824 DAG.getConstant(Intrinsic::aarch64_neon_uaddlv, DL, MVT::i32), CtPop); 6825 6826 if (VT == MVT::i64) 6827 UaddLV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, UaddLV); 6828 return UaddLV; 6829 } else if (VT == MVT::i128) { 6830 Val = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Val); 6831 6832 SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v16i8, Val); 6833 SDValue UaddLV = DAG.getNode( 6834 ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32, 6835 DAG.getConstant(Intrinsic::aarch64_neon_uaddlv, DL, MVT::i32), CtPop); 6836 6837 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i128, UaddLV); 6838 } 6839 6840 if (VT.isScalableVector() || useSVEForFixedLengthVectorVT(VT)) 6841 return LowerToPredicatedOp(Op, DAG, AArch64ISD::CTPOP_MERGE_PASSTHRU); 6842 6843 assert((VT == MVT::v1i64 || VT == MVT::v2i64 || VT == MVT::v2i32 || 6844 VT == MVT::v4i32 || VT == MVT::v4i16 || VT == MVT::v8i16) && 6845 "Unexpected type for custom ctpop lowering"); 6846 6847 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8; 6848 Val = DAG.getBitcast(VT8Bit, Val); 6849 Val = DAG.getNode(ISD::CTPOP, DL, VT8Bit, Val); 6850 6851 // Widen v8i8/v16i8 CTPOP result to VT by repeatedly widening pairwise adds. 6852 unsigned EltSize = 8; 6853 unsigned NumElts = VT.is64BitVector() ? 8 : 16; 6854 while (EltSize != VT.getScalarSizeInBits()) { 6855 EltSize *= 2; 6856 NumElts /= 2; 6857 MVT WidenVT = MVT::getVectorVT(MVT::getIntegerVT(EltSize), NumElts); 6858 Val = DAG.getNode( 6859 ISD::INTRINSIC_WO_CHAIN, DL, WidenVT, 6860 DAG.getConstant(Intrinsic::aarch64_neon_uaddlp, DL, MVT::i32), Val); 6861 } 6862 6863 return Val; 6864 } 6865 6866 SDValue AArch64TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const { 6867 EVT VT = Op.getValueType(); 6868 assert(VT.isScalableVector() || 6869 useSVEForFixedLengthVectorVT(VT, /*OverrideNEON=*/true)); 6870 6871 SDLoc DL(Op); 6872 SDValue RBIT = DAG.getNode(ISD::BITREVERSE, DL, VT, Op.getOperand(0)); 6873 return DAG.getNode(ISD::CTLZ, DL, VT, RBIT); 6874 } 6875 6876 SDValue AArch64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const { 6877 6878 if (Op.getValueType().isVector()) 6879 return LowerVSETCC(Op, DAG); 6880 6881 bool IsStrict = Op->isStrictFPOpcode(); 6882 bool IsSignaling = Op.getOpcode() == ISD::STRICT_FSETCCS; 6883 unsigned OpNo = IsStrict ? 1 : 0; 6884 SDValue Chain; 6885 if (IsStrict) 6886 Chain = Op.getOperand(0); 6887 SDValue LHS = Op.getOperand(OpNo + 0); 6888 SDValue RHS = Op.getOperand(OpNo + 1); 6889 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(OpNo + 2))->get(); 6890 SDLoc dl(Op); 6891 6892 // We chose ZeroOrOneBooleanContents, so use zero and one. 6893 EVT VT = Op.getValueType(); 6894 SDValue TVal = DAG.getConstant(1, dl, VT); 6895 SDValue FVal = DAG.getConstant(0, dl, VT); 6896 6897 // Handle f128 first, since one possible outcome is a normal integer 6898 // comparison which gets picked up by the next if statement. 6899 if (LHS.getValueType() == MVT::f128) { 6900 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl, LHS, RHS, Chain, 6901 IsSignaling); 6902 6903 // If softenSetCCOperands returned a scalar, use it. 6904 if (!RHS.getNode()) { 6905 assert(LHS.getValueType() == Op.getValueType() && 6906 "Unexpected setcc expansion!"); 6907 return IsStrict ? DAG.getMergeValues({LHS, Chain}, dl) : LHS; 6908 } 6909 } 6910 6911 if (LHS.getValueType().isInteger()) { 6912 SDValue CCVal; 6913 SDValue Cmp = getAArch64Cmp( 6914 LHS, RHS, ISD::getSetCCInverse(CC, LHS.getValueType()), CCVal, DAG, dl); 6915 6916 // Note that we inverted the condition above, so we reverse the order of 6917 // the true and false operands here. This will allow the setcc to be 6918 // matched to a single CSINC instruction. 6919 SDValue Res = DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CCVal, Cmp); 6920 return IsStrict ? DAG.getMergeValues({Res, Chain}, dl) : Res; 6921 } 6922 6923 // Now we know we're dealing with FP values. 6924 assert(LHS.getValueType() == MVT::f16 || LHS.getValueType() == MVT::f32 || 6925 LHS.getValueType() == MVT::f64); 6926 6927 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead 6928 // and do the comparison. 6929 SDValue Cmp; 6930 if (IsStrict) 6931 Cmp = emitStrictFPComparison(LHS, RHS, dl, DAG, Chain, IsSignaling); 6932 else 6933 Cmp = emitComparison(LHS, RHS, CC, dl, DAG); 6934 6935 AArch64CC::CondCode CC1, CC2; 6936 changeFPCCToAArch64CC(CC, CC1, CC2); 6937 SDValue Res; 6938 if (CC2 == AArch64CC::AL) { 6939 changeFPCCToAArch64CC(ISD::getSetCCInverse(CC, LHS.getValueType()), CC1, 6940 CC2); 6941 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32); 6942 6943 // Note that we inverted the condition above, so we reverse the order of 6944 // the true and false operands here. This will allow the setcc to be 6945 // matched to a single CSINC instruction. 6946 Res = DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CC1Val, Cmp); 6947 } else { 6948 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't 6949 // totally clean. Some of them require two CSELs to implement. As is in 6950 // this case, we emit the first CSEL and then emit a second using the output 6951 // of the first as the RHS. We're effectively OR'ing the two CC's together. 6952 6953 // FIXME: It would be nice if we could match the two CSELs to two CSINCs. 6954 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32); 6955 SDValue CS1 = 6956 DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp); 6957 6958 SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32); 6959 Res = DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp); 6960 } 6961 return IsStrict ? DAG.getMergeValues({Res, Cmp.getValue(1)}, dl) : Res; 6962 } 6963 6964 SDValue AArch64TargetLowering::LowerSELECT_CC(ISD::CondCode CC, SDValue LHS, 6965 SDValue RHS, SDValue TVal, 6966 SDValue FVal, const SDLoc &dl, 6967 SelectionDAG &DAG) const { 6968 // Handle f128 first, because it will result in a comparison of some RTLIB 6969 // call result against zero. 6970 if (LHS.getValueType() == MVT::f128) { 6971 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl, LHS, RHS); 6972 6973 // If softenSetCCOperands returned a scalar, we need to compare the result 6974 // against zero to select between true and false values. 6975 if (!RHS.getNode()) { 6976 RHS = DAG.getConstant(0, dl, LHS.getValueType()); 6977 CC = ISD::SETNE; 6978 } 6979 } 6980 6981 // Also handle f16, for which we need to do a f32 comparison. 6982 if (LHS.getValueType() == MVT::f16 && !Subtarget->hasFullFP16()) { 6983 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, LHS); 6984 RHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, RHS); 6985 } 6986 6987 // Next, handle integers. 6988 if (LHS.getValueType().isInteger()) { 6989 assert((LHS.getValueType() == RHS.getValueType()) && 6990 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64)); 6991 6992 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal); 6993 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal); 6994 ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS); 6995 // Check for sign pattern (SELECT_CC setgt, iN lhs, -1, 1, -1) and transform 6996 // into (OR (ASR lhs, N-1), 1), which requires less instructions for the 6997 // supported types. 6998 if (CC == ISD::SETGT && RHSC && RHSC->isAllOnesValue() && CTVal && CFVal && 6999 CTVal->isOne() && CFVal->isAllOnesValue() && 7000 LHS.getValueType() == TVal.getValueType()) { 7001 EVT VT = LHS.getValueType(); 7002 SDValue Shift = 7003 DAG.getNode(ISD::SRA, dl, VT, LHS, 7004 DAG.getConstant(VT.getSizeInBits() - 1, dl, VT)); 7005 return DAG.getNode(ISD::OR, dl, VT, Shift, DAG.getConstant(1, dl, VT)); 7006 } 7007 7008 unsigned Opcode = AArch64ISD::CSEL; 7009 7010 // If both the TVal and the FVal are constants, see if we can swap them in 7011 // order to for a CSINV or CSINC out of them. 7012 if (CTVal && CFVal && CTVal->isAllOnesValue() && CFVal->isNullValue()) { 7013 std::swap(TVal, FVal); 7014 std::swap(CTVal, CFVal); 7015 CC = ISD::getSetCCInverse(CC, LHS.getValueType()); 7016 } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isNullValue()) { 7017 std::swap(TVal, FVal); 7018 std::swap(CTVal, CFVal); 7019 CC = ISD::getSetCCInverse(CC, LHS.getValueType()); 7020 } else if (TVal.getOpcode() == ISD::XOR) { 7021 // If TVal is a NOT we want to swap TVal and FVal so that we can match 7022 // with a CSINV rather than a CSEL. 7023 if (isAllOnesConstant(TVal.getOperand(1))) { 7024 std::swap(TVal, FVal); 7025 std::swap(CTVal, CFVal); 7026 CC = ISD::getSetCCInverse(CC, LHS.getValueType()); 7027 } 7028 } else if (TVal.getOpcode() == ISD::SUB) { 7029 // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so 7030 // that we can match with a CSNEG rather than a CSEL. 7031 if (isNullConstant(TVal.getOperand(0))) { 7032 std::swap(TVal, FVal); 7033 std::swap(CTVal, CFVal); 7034 CC = ISD::getSetCCInverse(CC, LHS.getValueType()); 7035 } 7036 } else if (CTVal && CFVal) { 7037 const int64_t TrueVal = CTVal->getSExtValue(); 7038 const int64_t FalseVal = CFVal->getSExtValue(); 7039 bool Swap = false; 7040 7041 // If both TVal and FVal are constants, see if FVal is the 7042 // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC 7043 // instead of a CSEL in that case. 7044 if (TrueVal == ~FalseVal) { 7045 Opcode = AArch64ISD::CSINV; 7046 } else if (FalseVal > std::numeric_limits<int64_t>::min() && 7047 TrueVal == -FalseVal) { 7048 Opcode = AArch64ISD::CSNEG; 7049 } else if (TVal.getValueType() == MVT::i32) { 7050 // If our operands are only 32-bit wide, make sure we use 32-bit 7051 // arithmetic for the check whether we can use CSINC. This ensures that 7052 // the addition in the check will wrap around properly in case there is 7053 // an overflow (which would not be the case if we do the check with 7054 // 64-bit arithmetic). 7055 const uint32_t TrueVal32 = CTVal->getZExtValue(); 7056 const uint32_t FalseVal32 = CFVal->getZExtValue(); 7057 7058 if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) { 7059 Opcode = AArch64ISD::CSINC; 7060 7061 if (TrueVal32 > FalseVal32) { 7062 Swap = true; 7063 } 7064 } 7065 // 64-bit check whether we can use CSINC. 7066 } else if ((TrueVal == FalseVal + 1) || (TrueVal + 1 == FalseVal)) { 7067 Opcode = AArch64ISD::CSINC; 7068 7069 if (TrueVal > FalseVal) { 7070 Swap = true; 7071 } 7072 } 7073 7074 // Swap TVal and FVal if necessary. 7075 if (Swap) { 7076 std::swap(TVal, FVal); 7077 std::swap(CTVal, CFVal); 7078 CC = ISD::getSetCCInverse(CC, LHS.getValueType()); 7079 } 7080 7081 if (Opcode != AArch64ISD::CSEL) { 7082 // Drop FVal since we can get its value by simply inverting/negating 7083 // TVal. 7084 FVal = TVal; 7085 } 7086 } 7087 7088 // Avoid materializing a constant when possible by reusing a known value in 7089 // a register. However, don't perform this optimization if the known value 7090 // is one, zero or negative one in the case of a CSEL. We can always 7091 // materialize these values using CSINC, CSEL and CSINV with wzr/xzr as the 7092 // FVal, respectively. 7093 ConstantSDNode *RHSVal = dyn_cast<ConstantSDNode>(RHS); 7094 if (Opcode == AArch64ISD::CSEL && RHSVal && !RHSVal->isOne() && 7095 !RHSVal->isNullValue() && !RHSVal->isAllOnesValue()) { 7096 AArch64CC::CondCode AArch64CC = changeIntCCToAArch64CC(CC); 7097 // Transform "a == C ? C : x" to "a == C ? a : x" and "a != C ? x : C" to 7098 // "a != C ? x : a" to avoid materializing C. 7099 if (CTVal && CTVal == RHSVal && AArch64CC == AArch64CC::EQ) 7100 TVal = LHS; 7101 else if (CFVal && CFVal == RHSVal && AArch64CC == AArch64CC::NE) 7102 FVal = LHS; 7103 } else if (Opcode == AArch64ISD::CSNEG && RHSVal && RHSVal->isOne()) { 7104 assert (CTVal && CFVal && "Expected constant operands for CSNEG."); 7105 // Use a CSINV to transform "a == C ? 1 : -1" to "a == C ? a : -1" to 7106 // avoid materializing C. 7107 AArch64CC::CondCode AArch64CC = changeIntCCToAArch64CC(CC); 7108 if (CTVal == RHSVal && AArch64CC == AArch64CC::EQ) { 7109 Opcode = AArch64ISD::CSINV; 7110 TVal = LHS; 7111 FVal = DAG.getConstant(0, dl, FVal.getValueType()); 7112 } 7113 } 7114 7115 SDValue CCVal; 7116 SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl); 7117 EVT VT = TVal.getValueType(); 7118 return DAG.getNode(Opcode, dl, VT, TVal, FVal, CCVal, Cmp); 7119 } 7120 7121 // Now we know we're dealing with FP values. 7122 assert(LHS.getValueType() == MVT::f16 || LHS.getValueType() == MVT::f32 || 7123 LHS.getValueType() == MVT::f64); 7124 assert(LHS.getValueType() == RHS.getValueType()); 7125 EVT VT = TVal.getValueType(); 7126 SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG); 7127 7128 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally 7129 // clean. Some of them require two CSELs to implement. 7130 AArch64CC::CondCode CC1, CC2; 7131 changeFPCCToAArch64CC(CC, CC1, CC2); 7132 7133 if (DAG.getTarget().Options.UnsafeFPMath) { 7134 // Transform "a == 0.0 ? 0.0 : x" to "a == 0.0 ? a : x" and 7135 // "a != 0.0 ? x : 0.0" to "a != 0.0 ? x : a" to avoid materializing 0.0. 7136 ConstantFPSDNode *RHSVal = dyn_cast<ConstantFPSDNode>(RHS); 7137 if (RHSVal && RHSVal->isZero()) { 7138 ConstantFPSDNode *CFVal = dyn_cast<ConstantFPSDNode>(FVal); 7139 ConstantFPSDNode *CTVal = dyn_cast<ConstantFPSDNode>(TVal); 7140 7141 if ((CC == ISD::SETEQ || CC == ISD::SETOEQ || CC == ISD::SETUEQ) && 7142 CTVal && CTVal->isZero() && TVal.getValueType() == LHS.getValueType()) 7143 TVal = LHS; 7144 else if ((CC == ISD::SETNE || CC == ISD::SETONE || CC == ISD::SETUNE) && 7145 CFVal && CFVal->isZero() && 7146 FVal.getValueType() == LHS.getValueType()) 7147 FVal = LHS; 7148 } 7149 } 7150 7151 // Emit first, and possibly only, CSEL. 7152 SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32); 7153 SDValue CS1 = DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp); 7154 7155 // If we need a second CSEL, emit it, using the output of the first as the 7156 // RHS. We're effectively OR'ing the two CC's together. 7157 if (CC2 != AArch64CC::AL) { 7158 SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32); 7159 return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp); 7160 } 7161 7162 // Otherwise, return the output of the first CSEL. 7163 return CS1; 7164 } 7165 7166 SDValue AArch64TargetLowering::LowerSELECT_CC(SDValue Op, 7167 SelectionDAG &DAG) const { 7168 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get(); 7169 SDValue LHS = Op.getOperand(0); 7170 SDValue RHS = Op.getOperand(1); 7171 SDValue TVal = Op.getOperand(2); 7172 SDValue FVal = Op.getOperand(3); 7173 SDLoc DL(Op); 7174 return LowerSELECT_CC(CC, LHS, RHS, TVal, FVal, DL, DAG); 7175 } 7176 7177 SDValue AArch64TargetLowering::LowerSELECT(SDValue Op, 7178 SelectionDAG &DAG) const { 7179 SDValue CCVal = Op->getOperand(0); 7180 SDValue TVal = Op->getOperand(1); 7181 SDValue FVal = Op->getOperand(2); 7182 SDLoc DL(Op); 7183 7184 EVT Ty = Op.getValueType(); 7185 if (Ty.isScalableVector()) { 7186 SDValue TruncCC = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, CCVal); 7187 MVT PredVT = MVT::getVectorVT(MVT::i1, Ty.getVectorElementCount()); 7188 SDValue SplatPred = DAG.getNode(ISD::SPLAT_VECTOR, DL, PredVT, TruncCC); 7189 return DAG.getNode(ISD::VSELECT, DL, Ty, SplatPred, TVal, FVal); 7190 } 7191 7192 if (useSVEForFixedLengthVectorVT(Ty)) { 7193 // FIXME: Ideally this would be the same as above using i1 types, however 7194 // for the moment we can't deal with fixed i1 vector types properly, so 7195 // instead extend the predicate to a result type sized integer vector. 7196 MVT SplatValVT = MVT::getIntegerVT(Ty.getScalarSizeInBits()); 7197 MVT PredVT = MVT::getVectorVT(SplatValVT, Ty.getVectorElementCount()); 7198 SDValue SplatVal = DAG.getSExtOrTrunc(CCVal, DL, SplatValVT); 7199 SDValue SplatPred = DAG.getNode(ISD::SPLAT_VECTOR, DL, PredVT, SplatVal); 7200 return DAG.getNode(ISD::VSELECT, DL, Ty, SplatPred, TVal, FVal); 7201 } 7202 7203 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select 7204 // instruction. 7205 if (ISD::isOverflowIntrOpRes(CCVal)) { 7206 // Only lower legal XALUO ops. 7207 if (!DAG.getTargetLoweringInfo().isTypeLegal(CCVal->getValueType(0))) 7208 return SDValue(); 7209 7210 AArch64CC::CondCode OFCC; 7211 SDValue Value, Overflow; 7212 std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, CCVal.getValue(0), DAG); 7213 SDValue CCVal = DAG.getConstant(OFCC, DL, MVT::i32); 7214 7215 return DAG.getNode(AArch64ISD::CSEL, DL, Op.getValueType(), TVal, FVal, 7216 CCVal, Overflow); 7217 } 7218 7219 // Lower it the same way as we would lower a SELECT_CC node. 7220 ISD::CondCode CC; 7221 SDValue LHS, RHS; 7222 if (CCVal.getOpcode() == ISD::SETCC) { 7223 LHS = CCVal.getOperand(0); 7224 RHS = CCVal.getOperand(1); 7225 CC = cast<CondCodeSDNode>(CCVal.getOperand(2))->get(); 7226 } else { 7227 LHS = CCVal; 7228 RHS = DAG.getConstant(0, DL, CCVal.getValueType()); 7229 CC = ISD::SETNE; 7230 } 7231 return LowerSELECT_CC(CC, LHS, RHS, TVal, FVal, DL, DAG); 7232 } 7233 7234 SDValue AArch64TargetLowering::LowerJumpTable(SDValue Op, 7235 SelectionDAG &DAG) const { 7236 // Jump table entries as PC relative offsets. No additional tweaking 7237 // is necessary here. Just get the address of the jump table. 7238 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op); 7239 7240 if (getTargetMachine().getCodeModel() == CodeModel::Large && 7241 !Subtarget->isTargetMachO()) { 7242 return getAddrLarge(JT, DAG); 7243 } else if (getTargetMachine().getCodeModel() == CodeModel::Tiny) { 7244 return getAddrTiny(JT, DAG); 7245 } 7246 return getAddr(JT, DAG); 7247 } 7248 7249 SDValue AArch64TargetLowering::LowerBR_JT(SDValue Op, 7250 SelectionDAG &DAG) const { 7251 // Jump table entries as PC relative offsets. No additional tweaking 7252 // is necessary here. Just get the address of the jump table. 7253 SDLoc DL(Op); 7254 SDValue JT = Op.getOperand(1); 7255 SDValue Entry = Op.getOperand(2); 7256 int JTI = cast<JumpTableSDNode>(JT.getNode())->getIndex(); 7257 7258 auto *AFI = DAG.getMachineFunction().getInfo<AArch64FunctionInfo>(); 7259 AFI->setJumpTableEntryInfo(JTI, 4, nullptr); 7260 7261 SDNode *Dest = 7262 DAG.getMachineNode(AArch64::JumpTableDest32, DL, MVT::i64, MVT::i64, JT, 7263 Entry, DAG.getTargetJumpTable(JTI, MVT::i32)); 7264 return DAG.getNode(ISD::BRIND, DL, MVT::Other, Op.getOperand(0), 7265 SDValue(Dest, 0)); 7266 } 7267 7268 SDValue AArch64TargetLowering::LowerConstantPool(SDValue Op, 7269 SelectionDAG &DAG) const { 7270 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op); 7271 7272 if (getTargetMachine().getCodeModel() == CodeModel::Large) { 7273 // Use the GOT for the large code model on iOS. 7274 if (Subtarget->isTargetMachO()) { 7275 return getGOT(CP, DAG); 7276 } 7277 return getAddrLarge(CP, DAG); 7278 } else if (getTargetMachine().getCodeModel() == CodeModel::Tiny) { 7279 return getAddrTiny(CP, DAG); 7280 } else { 7281 return getAddr(CP, DAG); 7282 } 7283 } 7284 7285 SDValue AArch64TargetLowering::LowerBlockAddress(SDValue Op, 7286 SelectionDAG &DAG) const { 7287 BlockAddressSDNode *BA = cast<BlockAddressSDNode>(Op); 7288 if (getTargetMachine().getCodeModel() == CodeModel::Large && 7289 !Subtarget->isTargetMachO()) { 7290 return getAddrLarge(BA, DAG); 7291 } else if (getTargetMachine().getCodeModel() == CodeModel::Tiny) { 7292 return getAddrTiny(BA, DAG); 7293 } 7294 return getAddr(BA, DAG); 7295 } 7296 7297 SDValue AArch64TargetLowering::LowerDarwin_VASTART(SDValue Op, 7298 SelectionDAG &DAG) const { 7299 AArch64FunctionInfo *FuncInfo = 7300 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>(); 7301 7302 SDLoc DL(Op); 7303 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), 7304 getPointerTy(DAG.getDataLayout())); 7305 FR = DAG.getZExtOrTrunc(FR, DL, getPointerMemTy(DAG.getDataLayout())); 7306 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); 7307 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1), 7308 MachinePointerInfo(SV)); 7309 } 7310 7311 SDValue AArch64TargetLowering::LowerWin64_VASTART(SDValue Op, 7312 SelectionDAG &DAG) const { 7313 AArch64FunctionInfo *FuncInfo = 7314 DAG.getMachineFunction().getInfo<AArch64FunctionInfo>(); 7315 7316 SDLoc DL(Op); 7317 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsGPRSize() > 0 7318 ? FuncInfo->getVarArgsGPRIndex() 7319 : FuncInfo->getVarArgsStackIndex(), 7320 getPointerTy(DAG.getDataLayout())); 7321 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); 7322 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1), 7323 MachinePointerInfo(SV)); 7324 } 7325 7326 SDValue AArch64TargetLowering::LowerAAPCS_VASTART(SDValue Op, 7327 SelectionDAG &DAG) const { 7328 // The layout of the va_list struct is specified in the AArch64 Procedure Call 7329 // Standard, section B.3. 7330 MachineFunction &MF = DAG.getMachineFunction(); 7331 AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>(); 7332 unsigned PtrSize = Subtarget->isTargetILP32() ? 4 : 8; 7333 auto PtrMemVT = getPointerMemTy(DAG.getDataLayout()); 7334 auto PtrVT = getPointerTy(DAG.getDataLayout()); 7335 SDLoc DL(Op); 7336 7337 SDValue Chain = Op.getOperand(0); 7338 SDValue VAList = Op.getOperand(1); 7339 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); 7340 SmallVector<SDValue, 4> MemOps; 7341 7342 // void *__stack at offset 0 7343 unsigned Offset = 0; 7344 SDValue Stack = DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), PtrVT); 7345 Stack = DAG.getZExtOrTrunc(Stack, DL, PtrMemVT); 7346 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList, 7347 MachinePointerInfo(SV), Align(PtrSize))); 7348 7349 // void *__gr_top at offset 8 (4 on ILP32) 7350 Offset += PtrSize; 7351 int GPRSize = FuncInfo->getVarArgsGPRSize(); 7352 if (GPRSize > 0) { 7353 SDValue GRTop, GRTopAddr; 7354 7355 GRTopAddr = DAG.getNode(ISD::ADD, DL, PtrVT, VAList, 7356 DAG.getConstant(Offset, DL, PtrVT)); 7357 7358 GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), PtrVT); 7359 GRTop = DAG.getNode(ISD::ADD, DL, PtrVT, GRTop, 7360 DAG.getConstant(GPRSize, DL, PtrVT)); 7361 GRTop = DAG.getZExtOrTrunc(GRTop, DL, PtrMemVT); 7362 7363 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr, 7364 MachinePointerInfo(SV, Offset), 7365 Align(PtrSize))); 7366 } 7367 7368 // void *__vr_top at offset 16 (8 on ILP32) 7369 Offset += PtrSize; 7370 int FPRSize = FuncInfo->getVarArgsFPRSize(); 7371 if (FPRSize > 0) { 7372 SDValue VRTop, VRTopAddr; 7373 VRTopAddr = DAG.getNode(ISD::ADD, DL, PtrVT, VAList, 7374 DAG.getConstant(Offset, DL, PtrVT)); 7375 7376 VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), PtrVT); 7377 VRTop = DAG.getNode(ISD::ADD, DL, PtrVT, VRTop, 7378 DAG.getConstant(FPRSize, DL, PtrVT)); 7379 VRTop = DAG.getZExtOrTrunc(VRTop, DL, PtrMemVT); 7380 7381 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr, 7382 MachinePointerInfo(SV, Offset), 7383 Align(PtrSize))); 7384 } 7385 7386 // int __gr_offs at offset 24 (12 on ILP32) 7387 Offset += PtrSize; 7388 SDValue GROffsAddr = DAG.getNode(ISD::ADD, DL, PtrVT, VAList, 7389 DAG.getConstant(Offset, DL, PtrVT)); 7390 MemOps.push_back( 7391 DAG.getStore(Chain, DL, DAG.getConstant(-GPRSize, DL, MVT::i32), 7392 GROffsAddr, MachinePointerInfo(SV, Offset), Align(4))); 7393 7394 // int __vr_offs at offset 28 (16 on ILP32) 7395 Offset += 4; 7396 SDValue VROffsAddr = DAG.getNode(ISD::ADD, DL, PtrVT, VAList, 7397 DAG.getConstant(Offset, DL, PtrVT)); 7398 MemOps.push_back( 7399 DAG.getStore(Chain, DL, DAG.getConstant(-FPRSize, DL, MVT::i32), 7400 VROffsAddr, MachinePointerInfo(SV, Offset), Align(4))); 7401 7402 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps); 7403 } 7404 7405 SDValue AArch64TargetLowering::LowerVASTART(SDValue Op, 7406 SelectionDAG &DAG) const { 7407 MachineFunction &MF = DAG.getMachineFunction(); 7408 7409 if (Subtarget->isCallingConvWin64(MF.getFunction().getCallingConv())) 7410 return LowerWin64_VASTART(Op, DAG); 7411 else if (Subtarget->isTargetDarwin()) 7412 return LowerDarwin_VASTART(Op, DAG); 7413 else 7414 return LowerAAPCS_VASTART(Op, DAG); 7415 } 7416 7417 SDValue AArch64TargetLowering::LowerVACOPY(SDValue Op, 7418 SelectionDAG &DAG) const { 7419 // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single 7420 // pointer. 7421 SDLoc DL(Op); 7422 unsigned PtrSize = Subtarget->isTargetILP32() ? 4 : 8; 7423 unsigned VaListSize = 7424 (Subtarget->isTargetDarwin() || Subtarget->isTargetWindows()) 7425 ? PtrSize 7426 : Subtarget->isTargetILP32() ? 20 : 32; 7427 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue(); 7428 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue(); 7429 7430 return DAG.getMemcpy(Op.getOperand(0), DL, Op.getOperand(1), Op.getOperand(2), 7431 DAG.getConstant(VaListSize, DL, MVT::i32), 7432 Align(PtrSize), false, false, false, 7433 MachinePointerInfo(DestSV), MachinePointerInfo(SrcSV)); 7434 } 7435 7436 SDValue AArch64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const { 7437 assert(Subtarget->isTargetDarwin() && 7438 "automatic va_arg instruction only works on Darwin"); 7439 7440 const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); 7441 EVT VT = Op.getValueType(); 7442 SDLoc DL(Op); 7443 SDValue Chain = Op.getOperand(0); 7444 SDValue Addr = Op.getOperand(1); 7445 MaybeAlign Align(Op.getConstantOperandVal(3)); 7446 unsigned MinSlotSize = Subtarget->isTargetILP32() ? 4 : 8; 7447 auto PtrVT = getPointerTy(DAG.getDataLayout()); 7448 auto PtrMemVT = getPointerMemTy(DAG.getDataLayout()); 7449 SDValue VAList = 7450 DAG.getLoad(PtrMemVT, DL, Chain, Addr, MachinePointerInfo(V)); 7451 Chain = VAList.getValue(1); 7452 VAList = DAG.getZExtOrTrunc(VAList, DL, PtrVT); 7453 7454 if (VT.isScalableVector()) 7455 report_fatal_error("Passing SVE types to variadic functions is " 7456 "currently not supported"); 7457 7458 if (Align && *Align > MinSlotSize) { 7459 VAList = DAG.getNode(ISD::ADD, DL, PtrVT, VAList, 7460 DAG.getConstant(Align->value() - 1, DL, PtrVT)); 7461 VAList = DAG.getNode(ISD::AND, DL, PtrVT, VAList, 7462 DAG.getConstant(-(int64_t)Align->value(), DL, PtrVT)); 7463 } 7464 7465 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext()); 7466 unsigned ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy); 7467 7468 // Scalar integer and FP values smaller than 64 bits are implicitly extended 7469 // up to 64 bits. At the very least, we have to increase the striding of the 7470 // vaargs list to match this, and for FP values we need to introduce 7471 // FP_ROUND nodes as well. 7472 if (VT.isInteger() && !VT.isVector()) 7473 ArgSize = std::max(ArgSize, MinSlotSize); 7474 bool NeedFPTrunc = false; 7475 if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) { 7476 ArgSize = 8; 7477 NeedFPTrunc = true; 7478 } 7479 7480 // Increment the pointer, VAList, to the next vaarg 7481 SDValue VANext = DAG.getNode(ISD::ADD, DL, PtrVT, VAList, 7482 DAG.getConstant(ArgSize, DL, PtrVT)); 7483 VANext = DAG.getZExtOrTrunc(VANext, DL, PtrMemVT); 7484 7485 // Store the incremented VAList to the legalized pointer 7486 SDValue APStore = 7487 DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V)); 7488 7489 // Load the actual argument out of the pointer VAList 7490 if (NeedFPTrunc) { 7491 // Load the value as an f64. 7492 SDValue WideFP = 7493 DAG.getLoad(MVT::f64, DL, APStore, VAList, MachinePointerInfo()); 7494 // Round the value down to an f32. 7495 SDValue NarrowFP = DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0), 7496 DAG.getIntPtrConstant(1, DL)); 7497 SDValue Ops[] = { NarrowFP, WideFP.getValue(1) }; 7498 // Merge the rounded value with the chain output of the load. 7499 return DAG.getMergeValues(Ops, DL); 7500 } 7501 7502 return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo()); 7503 } 7504 7505 SDValue AArch64TargetLowering::LowerFRAMEADDR(SDValue Op, 7506 SelectionDAG &DAG) const { 7507 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); 7508 MFI.setFrameAddressIsTaken(true); 7509 7510 EVT VT = Op.getValueType(); 7511 SDLoc DL(Op); 7512 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 7513 SDValue FrameAddr = 7514 DAG.getCopyFromReg(DAG.getEntryNode(), DL, AArch64::FP, MVT::i64); 7515 while (Depth--) 7516 FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr, 7517 MachinePointerInfo()); 7518 7519 if (Subtarget->isTargetILP32()) 7520 FrameAddr = DAG.getNode(ISD::AssertZext, DL, MVT::i64, FrameAddr, 7521 DAG.getValueType(VT)); 7522 7523 return FrameAddr; 7524 } 7525 7526 SDValue AArch64TargetLowering::LowerSPONENTRY(SDValue Op, 7527 SelectionDAG &DAG) const { 7528 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); 7529 7530 EVT VT = getPointerTy(DAG.getDataLayout()); 7531 SDLoc DL(Op); 7532 int FI = MFI.CreateFixedObject(4, 0, false); 7533 return DAG.getFrameIndex(FI, VT); 7534 } 7535 7536 #define GET_REGISTER_MATCHER 7537 #include "AArch64GenAsmMatcher.inc" 7538 7539 // FIXME? Maybe this could be a TableGen attribute on some registers and 7540 // this table could be generated automatically from RegInfo. 7541 Register AArch64TargetLowering:: 7542 getRegisterByName(const char* RegName, LLT VT, const MachineFunction &MF) const { 7543 Register Reg = MatchRegisterName(RegName); 7544 if (AArch64::X1 <= Reg && Reg <= AArch64::X28) { 7545 const MCRegisterInfo *MRI = Subtarget->getRegisterInfo(); 7546 unsigned DwarfRegNum = MRI->getDwarfRegNum(Reg, false); 7547 if (!Subtarget->isXRegisterReserved(DwarfRegNum)) 7548 Reg = 0; 7549 } 7550 if (Reg) 7551 return Reg; 7552 report_fatal_error(Twine("Invalid register name \"" 7553 + StringRef(RegName) + "\".")); 7554 } 7555 7556 SDValue AArch64TargetLowering::LowerADDROFRETURNADDR(SDValue Op, 7557 SelectionDAG &DAG) const { 7558 DAG.getMachineFunction().getFrameInfo().setFrameAddressIsTaken(true); 7559 7560 EVT VT = Op.getValueType(); 7561 SDLoc DL(Op); 7562 7563 SDValue FrameAddr = 7564 DAG.getCopyFromReg(DAG.getEntryNode(), DL, AArch64::FP, VT); 7565 SDValue Offset = DAG.getConstant(8, DL, getPointerTy(DAG.getDataLayout())); 7566 7567 return DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset); 7568 } 7569 7570 SDValue AArch64TargetLowering::LowerRETURNADDR(SDValue Op, 7571 SelectionDAG &DAG) const { 7572 MachineFunction &MF = DAG.getMachineFunction(); 7573 MachineFrameInfo &MFI = MF.getFrameInfo(); 7574 MFI.setReturnAddressIsTaken(true); 7575 7576 EVT VT = Op.getValueType(); 7577 SDLoc DL(Op); 7578 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 7579 SDValue ReturnAddress; 7580 if (Depth) { 7581 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG); 7582 SDValue Offset = DAG.getConstant(8, DL, getPointerTy(DAG.getDataLayout())); 7583 ReturnAddress = DAG.getLoad( 7584 VT, DL, DAG.getEntryNode(), 7585 DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset), MachinePointerInfo()); 7586 } else { 7587 // Return LR, which contains the return address. Mark it an implicit 7588 // live-in. 7589 unsigned Reg = MF.addLiveIn(AArch64::LR, &AArch64::GPR64RegClass); 7590 ReturnAddress = DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT); 7591 } 7592 7593 // The XPACLRI instruction assembles to a hint-space instruction before 7594 // Armv8.3-A therefore this instruction can be safely used for any pre 7595 // Armv8.3-A architectures. On Armv8.3-A and onwards XPACI is available so use 7596 // that instead. 7597 SDNode *St; 7598 if (Subtarget->hasPAuth()) { 7599 St = DAG.getMachineNode(AArch64::XPACI, DL, VT, ReturnAddress); 7600 } else { 7601 // XPACLRI operates on LR therefore we must move the operand accordingly. 7602 SDValue Chain = 7603 DAG.getCopyToReg(DAG.getEntryNode(), DL, AArch64::LR, ReturnAddress); 7604 St = DAG.getMachineNode(AArch64::XPACLRI, DL, VT, Chain); 7605 } 7606 return SDValue(St, 0); 7607 } 7608 7609 /// LowerShiftParts - Lower SHL_PARTS/SRA_PARTS/SRL_PARTS, which returns two 7610 /// i32 values and take a 2 x i32 value to shift plus a shift amount. 7611 SDValue AArch64TargetLowering::LowerShiftParts(SDValue Op, 7612 SelectionDAG &DAG) const { 7613 SDValue Lo, Hi; 7614 expandShiftParts(Op.getNode(), Lo, Hi, DAG); 7615 return DAG.getMergeValues({Lo, Hi}, SDLoc(Op)); 7616 } 7617 7618 bool AArch64TargetLowering::isOffsetFoldingLegal( 7619 const GlobalAddressSDNode *GA) const { 7620 // Offsets are folded in the DAG combine rather than here so that we can 7621 // intelligently choose an offset based on the uses. 7622 return false; 7623 } 7624 7625 bool AArch64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT, 7626 bool OptForSize) const { 7627 bool IsLegal = false; 7628 // We can materialize #0.0 as fmov $Rd, XZR for 64-bit, 32-bit cases, and 7629 // 16-bit case when target has full fp16 support. 7630 // FIXME: We should be able to handle f128 as well with a clever lowering. 7631 const APInt ImmInt = Imm.bitcastToAPInt(); 7632 if (VT == MVT::f64) 7633 IsLegal = AArch64_AM::getFP64Imm(ImmInt) != -1 || Imm.isPosZero(); 7634 else if (VT == MVT::f32) 7635 IsLegal = AArch64_AM::getFP32Imm(ImmInt) != -1 || Imm.isPosZero(); 7636 else if (VT == MVT::f16 && Subtarget->hasFullFP16()) 7637 IsLegal = AArch64_AM::getFP16Imm(ImmInt) != -1 || Imm.isPosZero(); 7638 // TODO: fmov h0, w0 is also legal, however on't have an isel pattern to 7639 // generate that fmov. 7640 7641 // If we can not materialize in immediate field for fmov, check if the 7642 // value can be encoded as the immediate operand of a logical instruction. 7643 // The immediate value will be created with either MOVZ, MOVN, or ORR. 7644 if (!IsLegal && (VT == MVT::f64 || VT == MVT::f32)) { 7645 // The cost is actually exactly the same for mov+fmov vs. adrp+ldr; 7646 // however the mov+fmov sequence is always better because of the reduced 7647 // cache pressure. The timings are still the same if you consider 7648 // movw+movk+fmov vs. adrp+ldr (it's one instruction longer, but the 7649 // movw+movk is fused). So we limit up to 2 instrdduction at most. 7650 SmallVector<AArch64_IMM::ImmInsnModel, 4> Insn; 7651 AArch64_IMM::expandMOVImm(ImmInt.getZExtValue(), VT.getSizeInBits(), 7652 Insn); 7653 unsigned Limit = (OptForSize ? 1 : (Subtarget->hasFuseLiterals() ? 5 : 2)); 7654 IsLegal = Insn.size() <= Limit; 7655 } 7656 7657 LLVM_DEBUG(dbgs() << (IsLegal ? "Legal " : "Illegal ") << VT.getEVTString() 7658 << " imm value: "; Imm.dump();); 7659 return IsLegal; 7660 } 7661 7662 //===----------------------------------------------------------------------===// 7663 // AArch64 Optimization Hooks 7664 //===----------------------------------------------------------------------===// 7665 7666 static SDValue getEstimate(const AArch64Subtarget *ST, unsigned Opcode, 7667 SDValue Operand, SelectionDAG &DAG, 7668 int &ExtraSteps) { 7669 EVT VT = Operand.getValueType(); 7670 if (ST->hasNEON() && 7671 (VT == MVT::f64 || VT == MVT::v1f64 || VT == MVT::v2f64 || 7672 VT == MVT::f32 || VT == MVT::v1f32 || 7673 VT == MVT::v2f32 || VT == MVT::v4f32)) { 7674 if (ExtraSteps == TargetLoweringBase::ReciprocalEstimate::Unspecified) 7675 // For the reciprocal estimates, convergence is quadratic, so the number 7676 // of digits is doubled after each iteration. In ARMv8, the accuracy of 7677 // the initial estimate is 2^-8. Thus the number of extra steps to refine 7678 // the result for float (23 mantissa bits) is 2 and for double (52 7679 // mantissa bits) is 3. 7680 ExtraSteps = VT.getScalarType() == MVT::f64 ? 3 : 2; 7681 7682 return DAG.getNode(Opcode, SDLoc(Operand), VT, Operand); 7683 } 7684 7685 return SDValue(); 7686 } 7687 7688 SDValue 7689 AArch64TargetLowering::getSqrtInputTest(SDValue Op, SelectionDAG &DAG, 7690 const DenormalMode &Mode) const { 7691 SDLoc DL(Op); 7692 EVT VT = Op.getValueType(); 7693 EVT CCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); 7694 SDValue FPZero = DAG.getConstantFP(0.0, DL, VT); 7695 return DAG.getSetCC(DL, CCVT, Op, FPZero, ISD::SETEQ); 7696 } 7697 7698 SDValue 7699 AArch64TargetLowering::getSqrtResultForDenormInput(SDValue Op, 7700 SelectionDAG &DAG) const { 7701 return Op; 7702 } 7703 7704 SDValue AArch64TargetLowering::getSqrtEstimate(SDValue Operand, 7705 SelectionDAG &DAG, int Enabled, 7706 int &ExtraSteps, 7707 bool &UseOneConst, 7708 bool Reciprocal) const { 7709 if (Enabled == ReciprocalEstimate::Enabled || 7710 (Enabled == ReciprocalEstimate::Unspecified && Subtarget->useRSqrt())) 7711 if (SDValue Estimate = getEstimate(Subtarget, AArch64ISD::FRSQRTE, Operand, 7712 DAG, ExtraSteps)) { 7713 SDLoc DL(Operand); 7714 EVT VT = Operand.getValueType(); 7715 7716 SDNodeFlags Flags; 7717 Flags.setAllowReassociation(true); 7718 7719 // Newton reciprocal square root iteration: E * 0.5 * (3 - X * E^2) 7720 // AArch64 reciprocal square root iteration instruction: 0.5 * (3 - M * N) 7721 for (int i = ExtraSteps; i > 0; --i) { 7722 SDValue Step = DAG.getNode(ISD::FMUL, DL, VT, Estimate, Estimate, 7723 Flags); 7724 Step = DAG.getNode(AArch64ISD::FRSQRTS, DL, VT, Operand, Step, Flags); 7725 Estimate = DAG.getNode(ISD::FMUL, DL, VT, Estimate, Step, Flags); 7726 } 7727 if (!Reciprocal) 7728 Estimate = DAG.getNode(ISD::FMUL, DL, VT, Operand, Estimate, Flags); 7729 7730 ExtraSteps = 0; 7731 return Estimate; 7732 } 7733 7734 return SDValue(); 7735 } 7736 7737 SDValue AArch64TargetLowering::getRecipEstimate(SDValue Operand, 7738 SelectionDAG &DAG, int Enabled, 7739 int &ExtraSteps) const { 7740 if (Enabled == ReciprocalEstimate::Enabled) 7741 if (SDValue Estimate = getEstimate(Subtarget, AArch64ISD::FRECPE, Operand, 7742 DAG, ExtraSteps)) { 7743 SDLoc DL(Operand); 7744 EVT VT = Operand.getValueType(); 7745 7746 SDNodeFlags Flags; 7747 Flags.setAllowReassociation(true); 7748 7749 // Newton reciprocal iteration: E * (2 - X * E) 7750 // AArch64 reciprocal iteration instruction: (2 - M * N) 7751 for (int i = ExtraSteps; i > 0; --i) { 7752 SDValue Step = DAG.getNode(AArch64ISD::FRECPS, DL, VT, Operand, 7753 Estimate, Flags); 7754 Estimate = DAG.getNode(ISD::FMUL, DL, VT, Estimate, Step, Flags); 7755 } 7756 7757 ExtraSteps = 0; 7758 return Estimate; 7759 } 7760 7761 return SDValue(); 7762 } 7763 7764 //===----------------------------------------------------------------------===// 7765 // AArch64 Inline Assembly Support 7766 //===----------------------------------------------------------------------===// 7767 7768 // Table of Constraints 7769 // TODO: This is the current set of constraints supported by ARM for the 7770 // compiler, not all of them may make sense. 7771 // 7772 // r - A general register 7773 // w - An FP/SIMD register of some size in the range v0-v31 7774 // x - An FP/SIMD register of some size in the range v0-v15 7775 // I - Constant that can be used with an ADD instruction 7776 // J - Constant that can be used with a SUB instruction 7777 // K - Constant that can be used with a 32-bit logical instruction 7778 // L - Constant that can be used with a 64-bit logical instruction 7779 // M - Constant that can be used as a 32-bit MOV immediate 7780 // N - Constant that can be used as a 64-bit MOV immediate 7781 // Q - A memory reference with base register and no offset 7782 // S - A symbolic address 7783 // Y - Floating point constant zero 7784 // Z - Integer constant zero 7785 // 7786 // Note that general register operands will be output using their 64-bit x 7787 // register name, whatever the size of the variable, unless the asm operand 7788 // is prefixed by the %w modifier. Floating-point and SIMD register operands 7789 // will be output with the v prefix unless prefixed by the %b, %h, %s, %d or 7790 // %q modifier. 7791 const char *AArch64TargetLowering::LowerXConstraint(EVT ConstraintVT) const { 7792 // At this point, we have to lower this constraint to something else, so we 7793 // lower it to an "r" or "w". However, by doing this we will force the result 7794 // to be in register, while the X constraint is much more permissive. 7795 // 7796 // Although we are correct (we are free to emit anything, without 7797 // constraints), we might break use cases that would expect us to be more 7798 // efficient and emit something else. 7799 if (!Subtarget->hasFPARMv8()) 7800 return "r"; 7801 7802 if (ConstraintVT.isFloatingPoint()) 7803 return "w"; 7804 7805 if (ConstraintVT.isVector() && 7806 (ConstraintVT.getSizeInBits() == 64 || 7807 ConstraintVT.getSizeInBits() == 128)) 7808 return "w"; 7809 7810 return "r"; 7811 } 7812 7813 enum PredicateConstraint { 7814 Upl, 7815 Upa, 7816 Invalid 7817 }; 7818 7819 static PredicateConstraint parsePredicateConstraint(StringRef Constraint) { 7820 PredicateConstraint P = PredicateConstraint::Invalid; 7821 if (Constraint == "Upa") 7822 P = PredicateConstraint::Upa; 7823 if (Constraint == "Upl") 7824 P = PredicateConstraint::Upl; 7825 return P; 7826 } 7827 7828 /// getConstraintType - Given a constraint letter, return the type of 7829 /// constraint it is for this target. 7830 AArch64TargetLowering::ConstraintType 7831 AArch64TargetLowering::getConstraintType(StringRef Constraint) const { 7832 if (Constraint.size() == 1) { 7833 switch (Constraint[0]) { 7834 default: 7835 break; 7836 case 'x': 7837 case 'w': 7838 case 'y': 7839 return C_RegisterClass; 7840 // An address with a single base register. Due to the way we 7841 // currently handle addresses it is the same as 'r'. 7842 case 'Q': 7843 return C_Memory; 7844 case 'I': 7845 case 'J': 7846 case 'K': 7847 case 'L': 7848 case 'M': 7849 case 'N': 7850 case 'Y': 7851 case 'Z': 7852 return C_Immediate; 7853 case 'z': 7854 case 'S': // A symbolic address 7855 return C_Other; 7856 } 7857 } else if (parsePredicateConstraint(Constraint) != 7858 PredicateConstraint::Invalid) 7859 return C_RegisterClass; 7860 return TargetLowering::getConstraintType(Constraint); 7861 } 7862 7863 /// Examine constraint type and operand type and determine a weight value. 7864 /// This object must already have been set up with the operand type 7865 /// and the current alternative constraint selected. 7866 TargetLowering::ConstraintWeight 7867 AArch64TargetLowering::getSingleConstraintMatchWeight( 7868 AsmOperandInfo &info, const char *constraint) const { 7869 ConstraintWeight weight = CW_Invalid; 7870 Value *CallOperandVal = info.CallOperandVal; 7871 // If we don't have a value, we can't do a match, 7872 // but allow it at the lowest weight. 7873 if (!CallOperandVal) 7874 return CW_Default; 7875 Type *type = CallOperandVal->getType(); 7876 // Look at the constraint type. 7877 switch (*constraint) { 7878 default: 7879 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint); 7880 break; 7881 case 'x': 7882 case 'w': 7883 case 'y': 7884 if (type->isFloatingPointTy() || type->isVectorTy()) 7885 weight = CW_Register; 7886 break; 7887 case 'z': 7888 weight = CW_Constant; 7889 break; 7890 case 'U': 7891 if (parsePredicateConstraint(constraint) != PredicateConstraint::Invalid) 7892 weight = CW_Register; 7893 break; 7894 } 7895 return weight; 7896 } 7897 7898 std::pair<unsigned, const TargetRegisterClass *> 7899 AArch64TargetLowering::getRegForInlineAsmConstraint( 7900 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const { 7901 if (Constraint.size() == 1) { 7902 switch (Constraint[0]) { 7903 case 'r': 7904 if (VT.isScalableVector()) 7905 return std::make_pair(0U, nullptr); 7906 if (VT.getFixedSizeInBits() == 64) 7907 return std::make_pair(0U, &AArch64::GPR64commonRegClass); 7908 return std::make_pair(0U, &AArch64::GPR32commonRegClass); 7909 case 'w': { 7910 if (!Subtarget->hasFPARMv8()) 7911 break; 7912 if (VT.isScalableVector()) { 7913 if (VT.getVectorElementType() != MVT::i1) 7914 return std::make_pair(0U, &AArch64::ZPRRegClass); 7915 return std::make_pair(0U, nullptr); 7916 } 7917 uint64_t VTSize = VT.getFixedSizeInBits(); 7918 if (VTSize == 16) 7919 return std::make_pair(0U, &AArch64::FPR16RegClass); 7920 if (VTSize == 32) 7921 return std::make_pair(0U, &AArch64::FPR32RegClass); 7922 if (VTSize == 64) 7923 return std::make_pair(0U, &AArch64::FPR64RegClass); 7924 if (VTSize == 128) 7925 return std::make_pair(0U, &AArch64::FPR128RegClass); 7926 break; 7927 } 7928 // The instructions that this constraint is designed for can 7929 // only take 128-bit registers so just use that regclass. 7930 case 'x': 7931 if (!Subtarget->hasFPARMv8()) 7932 break; 7933 if (VT.isScalableVector()) 7934 return std::make_pair(0U, &AArch64::ZPR_4bRegClass); 7935 if (VT.getSizeInBits() == 128) 7936 return std::make_pair(0U, &AArch64::FPR128_loRegClass); 7937 break; 7938 case 'y': 7939 if (!Subtarget->hasFPARMv8()) 7940 break; 7941 if (VT.isScalableVector()) 7942 return std::make_pair(0U, &AArch64::ZPR_3bRegClass); 7943 break; 7944 } 7945 } else { 7946 PredicateConstraint PC = parsePredicateConstraint(Constraint); 7947 if (PC != PredicateConstraint::Invalid) { 7948 if (!VT.isScalableVector() || VT.getVectorElementType() != MVT::i1) 7949 return std::make_pair(0U, nullptr); 7950 bool restricted = (PC == PredicateConstraint::Upl); 7951 return restricted ? std::make_pair(0U, &AArch64::PPR_3bRegClass) 7952 : std::make_pair(0U, &AArch64::PPRRegClass); 7953 } 7954 } 7955 if (StringRef("{cc}").equals_lower(Constraint)) 7956 return std::make_pair(unsigned(AArch64::NZCV), &AArch64::CCRRegClass); 7957 7958 // Use the default implementation in TargetLowering to convert the register 7959 // constraint into a member of a register class. 7960 std::pair<unsigned, const TargetRegisterClass *> Res; 7961 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT); 7962 7963 // Not found as a standard register? 7964 if (!Res.second) { 7965 unsigned Size = Constraint.size(); 7966 if ((Size == 4 || Size == 5) && Constraint[0] == '{' && 7967 tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') { 7968 int RegNo; 7969 bool Failed = Constraint.slice(2, Size - 1).getAsInteger(10, RegNo); 7970 if (!Failed && RegNo >= 0 && RegNo <= 31) { 7971 // v0 - v31 are aliases of q0 - q31 or d0 - d31 depending on size. 7972 // By default we'll emit v0-v31 for this unless there's a modifier where 7973 // we'll emit the correct register as well. 7974 if (VT != MVT::Other && VT.getSizeInBits() == 64) { 7975 Res.first = AArch64::FPR64RegClass.getRegister(RegNo); 7976 Res.second = &AArch64::FPR64RegClass; 7977 } else { 7978 Res.first = AArch64::FPR128RegClass.getRegister(RegNo); 7979 Res.second = &AArch64::FPR128RegClass; 7980 } 7981 } 7982 } 7983 } 7984 7985 if (Res.second && !Subtarget->hasFPARMv8() && 7986 !AArch64::GPR32allRegClass.hasSubClassEq(Res.second) && 7987 !AArch64::GPR64allRegClass.hasSubClassEq(Res.second)) 7988 return std::make_pair(0U, nullptr); 7989 7990 return Res; 7991 } 7992 7993 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 7994 /// vector. If it is invalid, don't add anything to Ops. 7995 void AArch64TargetLowering::LowerAsmOperandForConstraint( 7996 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops, 7997 SelectionDAG &DAG) const { 7998 SDValue Result; 7999 8000 // Currently only support length 1 constraints. 8001 if (Constraint.length() != 1) 8002 return; 8003 8004 char ConstraintLetter = Constraint[0]; 8005 switch (ConstraintLetter) { 8006 default: 8007 break; 8008 8009 // This set of constraints deal with valid constants for various instructions. 8010 // Validate and return a target constant for them if we can. 8011 case 'z': { 8012 // 'z' maps to xzr or wzr so it needs an input of 0. 8013 if (!isNullConstant(Op)) 8014 return; 8015 8016 if (Op.getValueType() == MVT::i64) 8017 Result = DAG.getRegister(AArch64::XZR, MVT::i64); 8018 else 8019 Result = DAG.getRegister(AArch64::WZR, MVT::i32); 8020 break; 8021 } 8022 case 'S': { 8023 // An absolute symbolic address or label reference. 8024 if (const GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op)) { 8025 Result = DAG.getTargetGlobalAddress(GA->getGlobal(), SDLoc(Op), 8026 GA->getValueType(0)); 8027 } else if (const BlockAddressSDNode *BA = 8028 dyn_cast<BlockAddressSDNode>(Op)) { 8029 Result = 8030 DAG.getTargetBlockAddress(BA->getBlockAddress(), BA->getValueType(0)); 8031 } else if (const ExternalSymbolSDNode *ES = 8032 dyn_cast<ExternalSymbolSDNode>(Op)) { 8033 Result = 8034 DAG.getTargetExternalSymbol(ES->getSymbol(), ES->getValueType(0)); 8035 } else 8036 return; 8037 break; 8038 } 8039 8040 case 'I': 8041 case 'J': 8042 case 'K': 8043 case 'L': 8044 case 'M': 8045 case 'N': 8046 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op); 8047 if (!C) 8048 return; 8049 8050 // Grab the value and do some validation. 8051 uint64_t CVal = C->getZExtValue(); 8052 switch (ConstraintLetter) { 8053 // The I constraint applies only to simple ADD or SUB immediate operands: 8054 // i.e. 0 to 4095 with optional shift by 12 8055 // The J constraint applies only to ADD or SUB immediates that would be 8056 // valid when negated, i.e. if [an add pattern] were to be output as a SUB 8057 // instruction [or vice versa], in other words -1 to -4095 with optional 8058 // left shift by 12. 8059 case 'I': 8060 if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal)) 8061 break; 8062 return; 8063 case 'J': { 8064 uint64_t NVal = -C->getSExtValue(); 8065 if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal)) { 8066 CVal = C->getSExtValue(); 8067 break; 8068 } 8069 return; 8070 } 8071 // The K and L constraints apply *only* to logical immediates, including 8072 // what used to be the MOVI alias for ORR (though the MOVI alias has now 8073 // been removed and MOV should be used). So these constraints have to 8074 // distinguish between bit patterns that are valid 32-bit or 64-bit 8075 // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but 8076 // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice 8077 // versa. 8078 case 'K': 8079 if (AArch64_AM::isLogicalImmediate(CVal, 32)) 8080 break; 8081 return; 8082 case 'L': 8083 if (AArch64_AM::isLogicalImmediate(CVal, 64)) 8084 break; 8085 return; 8086 // The M and N constraints are a superset of K and L respectively, for use 8087 // with the MOV (immediate) alias. As well as the logical immediates they 8088 // also match 32 or 64-bit immediates that can be loaded either using a 8089 // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca 8090 // (M) or 64-bit 0x1234000000000000 (N) etc. 8091 // As a note some of this code is liberally stolen from the asm parser. 8092 case 'M': { 8093 if (!isUInt<32>(CVal)) 8094 return; 8095 if (AArch64_AM::isLogicalImmediate(CVal, 32)) 8096 break; 8097 if ((CVal & 0xFFFF) == CVal) 8098 break; 8099 if ((CVal & 0xFFFF0000ULL) == CVal) 8100 break; 8101 uint64_t NCVal = ~(uint32_t)CVal; 8102 if ((NCVal & 0xFFFFULL) == NCVal) 8103 break; 8104 if ((NCVal & 0xFFFF0000ULL) == NCVal) 8105 break; 8106 return; 8107 } 8108 case 'N': { 8109 if (AArch64_AM::isLogicalImmediate(CVal, 64)) 8110 break; 8111 if ((CVal & 0xFFFFULL) == CVal) 8112 break; 8113 if ((CVal & 0xFFFF0000ULL) == CVal) 8114 break; 8115 if ((CVal & 0xFFFF00000000ULL) == CVal) 8116 break; 8117 if ((CVal & 0xFFFF000000000000ULL) == CVal) 8118 break; 8119 uint64_t NCVal = ~CVal; 8120 if ((NCVal & 0xFFFFULL) == NCVal) 8121 break; 8122 if ((NCVal & 0xFFFF0000ULL) == NCVal) 8123 break; 8124 if ((NCVal & 0xFFFF00000000ULL) == NCVal) 8125 break; 8126 if ((NCVal & 0xFFFF000000000000ULL) == NCVal) 8127 break; 8128 return; 8129 } 8130 default: 8131 return; 8132 } 8133 8134 // All assembler immediates are 64-bit integers. 8135 Result = DAG.getTargetConstant(CVal, SDLoc(Op), MVT::i64); 8136 break; 8137 } 8138 8139 if (Result.getNode()) { 8140 Ops.push_back(Result); 8141 return; 8142 } 8143 8144 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); 8145 } 8146 8147 //===----------------------------------------------------------------------===// 8148 // AArch64 Advanced SIMD Support 8149 //===----------------------------------------------------------------------===// 8150 8151 /// WidenVector - Given a value in the V64 register class, produce the 8152 /// equivalent value in the V128 register class. 8153 static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) { 8154 EVT VT = V64Reg.getValueType(); 8155 unsigned NarrowSize = VT.getVectorNumElements(); 8156 MVT EltTy = VT.getVectorElementType().getSimpleVT(); 8157 MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize); 8158 SDLoc DL(V64Reg); 8159 8160 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy), 8161 V64Reg, DAG.getConstant(0, DL, MVT::i64)); 8162 } 8163 8164 /// getExtFactor - Determine the adjustment factor for the position when 8165 /// generating an "extract from vector registers" instruction. 8166 static unsigned getExtFactor(SDValue &V) { 8167 EVT EltType = V.getValueType().getVectorElementType(); 8168 return EltType.getSizeInBits() / 8; 8169 } 8170 8171 /// NarrowVector - Given a value in the V128 register class, produce the 8172 /// equivalent value in the V64 register class. 8173 static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) { 8174 EVT VT = V128Reg.getValueType(); 8175 unsigned WideSize = VT.getVectorNumElements(); 8176 MVT EltTy = VT.getVectorElementType().getSimpleVT(); 8177 MVT NarrowTy = MVT::getVectorVT(EltTy, WideSize / 2); 8178 SDLoc DL(V128Reg); 8179 8180 return DAG.getTargetExtractSubreg(AArch64::dsub, DL, NarrowTy, V128Reg); 8181 } 8182 8183 // Gather data to see if the operation can be modelled as a 8184 // shuffle in combination with VEXTs. 8185 SDValue AArch64TargetLowering::ReconstructShuffle(SDValue Op, 8186 SelectionDAG &DAG) const { 8187 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!"); 8188 LLVM_DEBUG(dbgs() << "AArch64TargetLowering::ReconstructShuffle\n"); 8189 SDLoc dl(Op); 8190 EVT VT = Op.getValueType(); 8191 assert(!VT.isScalableVector() && 8192 "Scalable vectors cannot be used with ISD::BUILD_VECTOR"); 8193 unsigned NumElts = VT.getVectorNumElements(); 8194 8195 struct ShuffleSourceInfo { 8196 SDValue Vec; 8197 unsigned MinElt; 8198 unsigned MaxElt; 8199 8200 // We may insert some combination of BITCASTs and VEXT nodes to force Vec to 8201 // be compatible with the shuffle we intend to construct. As a result 8202 // ShuffleVec will be some sliding window into the original Vec. 8203 SDValue ShuffleVec; 8204 8205 // Code should guarantee that element i in Vec starts at element "WindowBase 8206 // + i * WindowScale in ShuffleVec". 8207 int WindowBase; 8208 int WindowScale; 8209 8210 ShuffleSourceInfo(SDValue Vec) 8211 : Vec(Vec), MinElt(std::numeric_limits<unsigned>::max()), MaxElt(0), 8212 ShuffleVec(Vec), WindowBase(0), WindowScale(1) {} 8213 8214 bool operator ==(SDValue OtherVec) { return Vec == OtherVec; } 8215 }; 8216 8217 // First gather all vectors used as an immediate source for this BUILD_VECTOR 8218 // node. 8219 SmallVector<ShuffleSourceInfo, 2> Sources; 8220 for (unsigned i = 0; i < NumElts; ++i) { 8221 SDValue V = Op.getOperand(i); 8222 if (V.isUndef()) 8223 continue; 8224 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT || 8225 !isa<ConstantSDNode>(V.getOperand(1))) { 8226 LLVM_DEBUG( 8227 dbgs() << "Reshuffle failed: " 8228 "a shuffle can only come from building a vector from " 8229 "various elements of other vectors, provided their " 8230 "indices are constant\n"); 8231 return SDValue(); 8232 } 8233 8234 // Add this element source to the list if it's not already there. 8235 SDValue SourceVec = V.getOperand(0); 8236 auto Source = find(Sources, SourceVec); 8237 if (Source == Sources.end()) 8238 Source = Sources.insert(Sources.end(), ShuffleSourceInfo(SourceVec)); 8239 8240 // Update the minimum and maximum lane number seen. 8241 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue(); 8242 Source->MinElt = std::min(Source->MinElt, EltNo); 8243 Source->MaxElt = std::max(Source->MaxElt, EltNo); 8244 } 8245 8246 if (Sources.size() > 2) { 8247 LLVM_DEBUG( 8248 dbgs() << "Reshuffle failed: currently only do something sane when at " 8249 "most two source vectors are involved\n"); 8250 return SDValue(); 8251 } 8252 8253 // Find out the smallest element size among result and two sources, and use 8254 // it as element size to build the shuffle_vector. 8255 EVT SmallestEltTy = VT.getVectorElementType(); 8256 for (auto &Source : Sources) { 8257 EVT SrcEltTy = Source.Vec.getValueType().getVectorElementType(); 8258 if (SrcEltTy.bitsLT(SmallestEltTy)) { 8259 SmallestEltTy = SrcEltTy; 8260 } 8261 } 8262 unsigned ResMultiplier = 8263 VT.getScalarSizeInBits() / SmallestEltTy.getFixedSizeInBits(); 8264 uint64_t VTSize = VT.getFixedSizeInBits(); 8265 NumElts = VTSize / SmallestEltTy.getFixedSizeInBits(); 8266 EVT ShuffleVT = EVT::getVectorVT(*DAG.getContext(), SmallestEltTy, NumElts); 8267 8268 // If the source vector is too wide or too narrow, we may nevertheless be able 8269 // to construct a compatible shuffle either by concatenating it with UNDEF or 8270 // extracting a suitable range of elements. 8271 for (auto &Src : Sources) { 8272 EVT SrcVT = Src.ShuffleVec.getValueType(); 8273 8274 uint64_t SrcVTSize = SrcVT.getFixedSizeInBits(); 8275 if (SrcVTSize == VTSize) 8276 continue; 8277 8278 // This stage of the search produces a source with the same element type as 8279 // the original, but with a total width matching the BUILD_VECTOR output. 8280 EVT EltVT = SrcVT.getVectorElementType(); 8281 unsigned NumSrcElts = VTSize / EltVT.getFixedSizeInBits(); 8282 EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumSrcElts); 8283 8284 if (SrcVTSize < VTSize) { 8285 assert(2 * SrcVTSize == VTSize); 8286 // We can pad out the smaller vector for free, so if it's part of a 8287 // shuffle... 8288 Src.ShuffleVec = 8289 DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, Src.ShuffleVec, 8290 DAG.getUNDEF(Src.ShuffleVec.getValueType())); 8291 continue; 8292 } 8293 8294 if (SrcVTSize != 2 * VTSize) { 8295 LLVM_DEBUG( 8296 dbgs() << "Reshuffle failed: result vector too small to extract\n"); 8297 return SDValue(); 8298 } 8299 8300 if (Src.MaxElt - Src.MinElt >= NumSrcElts) { 8301 LLVM_DEBUG( 8302 dbgs() << "Reshuffle failed: span too large for a VEXT to cope\n"); 8303 return SDValue(); 8304 } 8305 8306 if (Src.MinElt >= NumSrcElts) { 8307 // The extraction can just take the second half 8308 Src.ShuffleVec = 8309 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec, 8310 DAG.getConstant(NumSrcElts, dl, MVT::i64)); 8311 Src.WindowBase = -NumSrcElts; 8312 } else if (Src.MaxElt < NumSrcElts) { 8313 // The extraction can just take the first half 8314 Src.ShuffleVec = 8315 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec, 8316 DAG.getConstant(0, dl, MVT::i64)); 8317 } else { 8318 // An actual VEXT is needed 8319 SDValue VEXTSrc1 = 8320 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec, 8321 DAG.getConstant(0, dl, MVT::i64)); 8322 SDValue VEXTSrc2 = 8323 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec, 8324 DAG.getConstant(NumSrcElts, dl, MVT::i64)); 8325 unsigned Imm = Src.MinElt * getExtFactor(VEXTSrc1); 8326 8327 if (!SrcVT.is64BitVector()) { 8328 LLVM_DEBUG( 8329 dbgs() << "Reshuffle failed: don't know how to lower AArch64ISD::EXT " 8330 "for SVE vectors."); 8331 return SDValue(); 8332 } 8333 8334 Src.ShuffleVec = DAG.getNode(AArch64ISD::EXT, dl, DestVT, VEXTSrc1, 8335 VEXTSrc2, 8336 DAG.getConstant(Imm, dl, MVT::i32)); 8337 Src.WindowBase = -Src.MinElt; 8338 } 8339 } 8340 8341 // Another possible incompatibility occurs from the vector element types. We 8342 // can fix this by bitcasting the source vectors to the same type we intend 8343 // for the shuffle. 8344 for (auto &Src : Sources) { 8345 EVT SrcEltTy = Src.ShuffleVec.getValueType().getVectorElementType(); 8346 if (SrcEltTy == SmallestEltTy) 8347 continue; 8348 assert(ShuffleVT.getVectorElementType() == SmallestEltTy); 8349 Src.ShuffleVec = DAG.getNode(ISD::BITCAST, dl, ShuffleVT, Src.ShuffleVec); 8350 Src.WindowScale = 8351 SrcEltTy.getFixedSizeInBits() / SmallestEltTy.getFixedSizeInBits(); 8352 Src.WindowBase *= Src.WindowScale; 8353 } 8354 8355 // Final sanity check before we try to actually produce a shuffle. 8356 LLVM_DEBUG(for (auto Src 8357 : Sources) 8358 assert(Src.ShuffleVec.getValueType() == ShuffleVT);); 8359 8360 // The stars all align, our next step is to produce the mask for the shuffle. 8361 SmallVector<int, 8> Mask(ShuffleVT.getVectorNumElements(), -1); 8362 int BitsPerShuffleLane = ShuffleVT.getScalarSizeInBits(); 8363 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) { 8364 SDValue Entry = Op.getOperand(i); 8365 if (Entry.isUndef()) 8366 continue; 8367 8368 auto Src = find(Sources, Entry.getOperand(0)); 8369 int EltNo = cast<ConstantSDNode>(Entry.getOperand(1))->getSExtValue(); 8370 8371 // EXTRACT_VECTOR_ELT performs an implicit any_ext; BUILD_VECTOR an implicit 8372 // trunc. So only std::min(SrcBits, DestBits) actually get defined in this 8373 // segment. 8374 EVT OrigEltTy = Entry.getOperand(0).getValueType().getVectorElementType(); 8375 int BitsDefined = std::min(OrigEltTy.getScalarSizeInBits(), 8376 VT.getScalarSizeInBits()); 8377 int LanesDefined = BitsDefined / BitsPerShuffleLane; 8378 8379 // This source is expected to fill ResMultiplier lanes of the final shuffle, 8380 // starting at the appropriate offset. 8381 int *LaneMask = &Mask[i * ResMultiplier]; 8382 8383 int ExtractBase = EltNo * Src->WindowScale + Src->WindowBase; 8384 ExtractBase += NumElts * (Src - Sources.begin()); 8385 for (int j = 0; j < LanesDefined; ++j) 8386 LaneMask[j] = ExtractBase + j; 8387 } 8388 8389 // Final check before we try to produce nonsense... 8390 if (!isShuffleMaskLegal(Mask, ShuffleVT)) { 8391 LLVM_DEBUG(dbgs() << "Reshuffle failed: illegal shuffle mask\n"); 8392 return SDValue(); 8393 } 8394 8395 SDValue ShuffleOps[] = { DAG.getUNDEF(ShuffleVT), DAG.getUNDEF(ShuffleVT) }; 8396 for (unsigned i = 0; i < Sources.size(); ++i) 8397 ShuffleOps[i] = Sources[i].ShuffleVec; 8398 8399 SDValue Shuffle = DAG.getVectorShuffle(ShuffleVT, dl, ShuffleOps[0], 8400 ShuffleOps[1], Mask); 8401 SDValue V = DAG.getNode(ISD::BITCAST, dl, VT, Shuffle); 8402 8403 LLVM_DEBUG(dbgs() << "Reshuffle, creating node: "; Shuffle.dump(); 8404 dbgs() << "Reshuffle, creating node: "; V.dump();); 8405 8406 return V; 8407 } 8408 8409 // check if an EXT instruction can handle the shuffle mask when the 8410 // vector sources of the shuffle are the same. 8411 static bool isSingletonEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) { 8412 unsigned NumElts = VT.getVectorNumElements(); 8413 8414 // Assume that the first shuffle index is not UNDEF. Fail if it is. 8415 if (M[0] < 0) 8416 return false; 8417 8418 Imm = M[0]; 8419 8420 // If this is a VEXT shuffle, the immediate value is the index of the first 8421 // element. The other shuffle indices must be the successive elements after 8422 // the first one. 8423 unsigned ExpectedElt = Imm; 8424 for (unsigned i = 1; i < NumElts; ++i) { 8425 // Increment the expected index. If it wraps around, just follow it 8426 // back to index zero and keep going. 8427 ++ExpectedElt; 8428 if (ExpectedElt == NumElts) 8429 ExpectedElt = 0; 8430 8431 if (M[i] < 0) 8432 continue; // ignore UNDEF indices 8433 if (ExpectedElt != static_cast<unsigned>(M[i])) 8434 return false; 8435 } 8436 8437 return true; 8438 } 8439 8440 /// Check if a vector shuffle corresponds to a DUP instructions with a larger 8441 /// element width than the vector lane type. If that is the case the function 8442 /// returns true and writes the value of the DUP instruction lane operand into 8443 /// DupLaneOp 8444 static bool isWideDUPMask(ArrayRef<int> M, EVT VT, unsigned BlockSize, 8445 unsigned &DupLaneOp) { 8446 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) && 8447 "Only possible block sizes for wide DUP are: 16, 32, 64"); 8448 8449 if (BlockSize <= VT.getScalarSizeInBits()) 8450 return false; 8451 if (BlockSize % VT.getScalarSizeInBits() != 0) 8452 return false; 8453 if (VT.getSizeInBits() % BlockSize != 0) 8454 return false; 8455 8456 size_t SingleVecNumElements = VT.getVectorNumElements(); 8457 size_t NumEltsPerBlock = BlockSize / VT.getScalarSizeInBits(); 8458 size_t NumBlocks = VT.getSizeInBits() / BlockSize; 8459 8460 // We are looking for masks like 8461 // [0, 1, 0, 1] or [2, 3, 2, 3] or [4, 5, 6, 7, 4, 5, 6, 7] where any element 8462 // might be replaced by 'undefined'. BlockIndices will eventually contain 8463 // lane indices of the duplicated block (i.e. [0, 1], [2, 3] and [4, 5, 6, 7] 8464 // for the above examples) 8465 SmallVector<int, 8> BlockElts(NumEltsPerBlock, -1); 8466 for (size_t BlockIndex = 0; BlockIndex < NumBlocks; BlockIndex++) 8467 for (size_t I = 0; I < NumEltsPerBlock; I++) { 8468 int Elt = M[BlockIndex * NumEltsPerBlock + I]; 8469 if (Elt < 0) 8470 continue; 8471 // For now we don't support shuffles that use the second operand 8472 if ((unsigned)Elt >= SingleVecNumElements) 8473 return false; 8474 if (BlockElts[I] < 0) 8475 BlockElts[I] = Elt; 8476 else if (BlockElts[I] != Elt) 8477 return false; 8478 } 8479 8480 // We found a candidate block (possibly with some undefs). It must be a 8481 // sequence of consecutive integers starting with a value divisible by 8482 // NumEltsPerBlock with some values possibly replaced by undef-s. 8483 8484 // Find first non-undef element 8485 auto FirstRealEltIter = find_if(BlockElts, [](int Elt) { return Elt >= 0; }); 8486 assert(FirstRealEltIter != BlockElts.end() && 8487 "Shuffle with all-undefs must have been caught by previous cases, " 8488 "e.g. isSplat()"); 8489 if (FirstRealEltIter == BlockElts.end()) { 8490 DupLaneOp = 0; 8491 return true; 8492 } 8493 8494 // Index of FirstRealElt in BlockElts 8495 size_t FirstRealIndex = FirstRealEltIter - BlockElts.begin(); 8496 8497 if ((unsigned)*FirstRealEltIter < FirstRealIndex) 8498 return false; 8499 // BlockElts[0] must have the following value if it isn't undef: 8500 size_t Elt0 = *FirstRealEltIter - FirstRealIndex; 8501 8502 // Check the first element 8503 if (Elt0 % NumEltsPerBlock != 0) 8504 return false; 8505 // Check that the sequence indeed consists of consecutive integers (modulo 8506 // undefs) 8507 for (size_t I = 0; I < NumEltsPerBlock; I++) 8508 if (BlockElts[I] >= 0 && (unsigned)BlockElts[I] != Elt0 + I) 8509 return false; 8510 8511 DupLaneOp = Elt0 / NumEltsPerBlock; 8512 return true; 8513 } 8514 8515 // check if an EXT instruction can handle the shuffle mask when the 8516 // vector sources of the shuffle are different. 8517 static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT, 8518 unsigned &Imm) { 8519 // Look for the first non-undef element. 8520 const int *FirstRealElt = find_if(M, [](int Elt) { return Elt >= 0; }); 8521 8522 // Benefit form APInt to handle overflow when calculating expected element. 8523 unsigned NumElts = VT.getVectorNumElements(); 8524 unsigned MaskBits = APInt(32, NumElts * 2).logBase2(); 8525 APInt ExpectedElt = APInt(MaskBits, *FirstRealElt + 1); 8526 // The following shuffle indices must be the successive elements after the 8527 // first real element. 8528 const int *FirstWrongElt = std::find_if(FirstRealElt + 1, M.end(), 8529 [&](int Elt) {return Elt != ExpectedElt++ && Elt != -1;}); 8530 if (FirstWrongElt != M.end()) 8531 return false; 8532 8533 // The index of an EXT is the first element if it is not UNDEF. 8534 // Watch out for the beginning UNDEFs. The EXT index should be the expected 8535 // value of the first element. E.g. 8536 // <-1, -1, 3, ...> is treated as <1, 2, 3, ...>. 8537 // <-1, -1, 0, 1, ...> is treated as <2*NumElts-2, 2*NumElts-1, 0, 1, ...>. 8538 // ExpectedElt is the last mask index plus 1. 8539 Imm = ExpectedElt.getZExtValue(); 8540 8541 // There are two difference cases requiring to reverse input vectors. 8542 // For example, for vector <4 x i32> we have the following cases, 8543 // Case 1: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, -1, 0>) 8544 // Case 2: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, 7, 0>) 8545 // For both cases, we finally use mask <5, 6, 7, 0>, which requires 8546 // to reverse two input vectors. 8547 if (Imm < NumElts) 8548 ReverseEXT = true; 8549 else 8550 Imm -= NumElts; 8551 8552 return true; 8553 } 8554 8555 /// isREVMask - Check if a vector shuffle corresponds to a REV 8556 /// instruction with the specified blocksize. (The order of the elements 8557 /// within each block of the vector is reversed.) 8558 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) { 8559 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) && 8560 "Only possible block sizes for REV are: 16, 32, 64"); 8561 8562 unsigned EltSz = VT.getScalarSizeInBits(); 8563 if (EltSz == 64) 8564 return false; 8565 8566 unsigned NumElts = VT.getVectorNumElements(); 8567 unsigned BlockElts = M[0] + 1; 8568 // If the first shuffle index is UNDEF, be optimistic. 8569 if (M[0] < 0) 8570 BlockElts = BlockSize / EltSz; 8571 8572 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz) 8573 return false; 8574 8575 for (unsigned i = 0; i < NumElts; ++i) { 8576 if (M[i] < 0) 8577 continue; // ignore UNDEF indices 8578 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts)) 8579 return false; 8580 } 8581 8582 return true; 8583 } 8584 8585 static bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) { 8586 unsigned NumElts = VT.getVectorNumElements(); 8587 if (NumElts % 2 != 0) 8588 return false; 8589 WhichResult = (M[0] == 0 ? 0 : 1); 8590 unsigned Idx = WhichResult * NumElts / 2; 8591 for (unsigned i = 0; i != NumElts; i += 2) { 8592 if ((M[i] >= 0 && (unsigned)M[i] != Idx) || 8593 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts)) 8594 return false; 8595 Idx += 1; 8596 } 8597 8598 return true; 8599 } 8600 8601 static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) { 8602 unsigned NumElts = VT.getVectorNumElements(); 8603 WhichResult = (M[0] == 0 ? 0 : 1); 8604 for (unsigned i = 0; i != NumElts; ++i) { 8605 if (M[i] < 0) 8606 continue; // ignore UNDEF indices 8607 if ((unsigned)M[i] != 2 * i + WhichResult) 8608 return false; 8609 } 8610 8611 return true; 8612 } 8613 8614 static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) { 8615 unsigned NumElts = VT.getVectorNumElements(); 8616 if (NumElts % 2 != 0) 8617 return false; 8618 WhichResult = (M[0] == 0 ? 0 : 1); 8619 for (unsigned i = 0; i < NumElts; i += 2) { 8620 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) || 8621 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + NumElts + WhichResult)) 8622 return false; 8623 } 8624 return true; 8625 } 8626 8627 /// isZIP_v_undef_Mask - Special case of isZIPMask for canonical form of 8628 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef". 8629 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>. 8630 static bool isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) { 8631 unsigned NumElts = VT.getVectorNumElements(); 8632 if (NumElts % 2 != 0) 8633 return false; 8634 WhichResult = (M[0] == 0 ? 0 : 1); 8635 unsigned Idx = WhichResult * NumElts / 2; 8636 for (unsigned i = 0; i != NumElts; i += 2) { 8637 if ((M[i] >= 0 && (unsigned)M[i] != Idx) || 8638 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx)) 8639 return false; 8640 Idx += 1; 8641 } 8642 8643 return true; 8644 } 8645 8646 /// isUZP_v_undef_Mask - Special case of isUZPMask for canonical form of 8647 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef". 8648 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>, 8649 static bool isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) { 8650 unsigned Half = VT.getVectorNumElements() / 2; 8651 WhichResult = (M[0] == 0 ? 0 : 1); 8652 for (unsigned j = 0; j != 2; ++j) { 8653 unsigned Idx = WhichResult; 8654 for (unsigned i = 0; i != Half; ++i) { 8655 int MIdx = M[i + j * Half]; 8656 if (MIdx >= 0 && (unsigned)MIdx != Idx) 8657 return false; 8658 Idx += 2; 8659 } 8660 } 8661 8662 return true; 8663 } 8664 8665 /// isTRN_v_undef_Mask - Special case of isTRNMask for canonical form of 8666 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef". 8667 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>. 8668 static bool isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) { 8669 unsigned NumElts = VT.getVectorNumElements(); 8670 if (NumElts % 2 != 0) 8671 return false; 8672 WhichResult = (M[0] == 0 ? 0 : 1); 8673 for (unsigned i = 0; i < NumElts; i += 2) { 8674 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) || 8675 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + WhichResult)) 8676 return false; 8677 } 8678 return true; 8679 } 8680 8681 static bool isINSMask(ArrayRef<int> M, int NumInputElements, 8682 bool &DstIsLeft, int &Anomaly) { 8683 if (M.size() != static_cast<size_t>(NumInputElements)) 8684 return false; 8685 8686 int NumLHSMatch = 0, NumRHSMatch = 0; 8687 int LastLHSMismatch = -1, LastRHSMismatch = -1; 8688 8689 for (int i = 0; i < NumInputElements; ++i) { 8690 if (M[i] == -1) { 8691 ++NumLHSMatch; 8692 ++NumRHSMatch; 8693 continue; 8694 } 8695 8696 if (M[i] == i) 8697 ++NumLHSMatch; 8698 else 8699 LastLHSMismatch = i; 8700 8701 if (M[i] == i + NumInputElements) 8702 ++NumRHSMatch; 8703 else 8704 LastRHSMismatch = i; 8705 } 8706 8707 if (NumLHSMatch == NumInputElements - 1) { 8708 DstIsLeft = true; 8709 Anomaly = LastLHSMismatch; 8710 return true; 8711 } else if (NumRHSMatch == NumInputElements - 1) { 8712 DstIsLeft = false; 8713 Anomaly = LastRHSMismatch; 8714 return true; 8715 } 8716 8717 return false; 8718 } 8719 8720 static bool isConcatMask(ArrayRef<int> Mask, EVT VT, bool SplitLHS) { 8721 if (VT.getSizeInBits() != 128) 8722 return false; 8723 8724 unsigned NumElts = VT.getVectorNumElements(); 8725 8726 for (int I = 0, E = NumElts / 2; I != E; I++) { 8727 if (Mask[I] != I) 8728 return false; 8729 } 8730 8731 int Offset = NumElts / 2; 8732 for (int I = NumElts / 2, E = NumElts; I != E; I++) { 8733 if (Mask[I] != I + SplitLHS * Offset) 8734 return false; 8735 } 8736 8737 return true; 8738 } 8739 8740 static SDValue tryFormConcatFromShuffle(SDValue Op, SelectionDAG &DAG) { 8741 SDLoc DL(Op); 8742 EVT VT = Op.getValueType(); 8743 SDValue V0 = Op.getOperand(0); 8744 SDValue V1 = Op.getOperand(1); 8745 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op)->getMask(); 8746 8747 if (VT.getVectorElementType() != V0.getValueType().getVectorElementType() || 8748 VT.getVectorElementType() != V1.getValueType().getVectorElementType()) 8749 return SDValue(); 8750 8751 bool SplitV0 = V0.getValueSizeInBits() == 128; 8752 8753 if (!isConcatMask(Mask, VT, SplitV0)) 8754 return SDValue(); 8755 8756 EVT CastVT = VT.getHalfNumVectorElementsVT(*DAG.getContext()); 8757 if (SplitV0) { 8758 V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V0, 8759 DAG.getConstant(0, DL, MVT::i64)); 8760 } 8761 if (V1.getValueSizeInBits() == 128) { 8762 V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V1, 8763 DAG.getConstant(0, DL, MVT::i64)); 8764 } 8765 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, V0, V1); 8766 } 8767 8768 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit 8769 /// the specified operations to build the shuffle. 8770 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS, 8771 SDValue RHS, SelectionDAG &DAG, 8772 const SDLoc &dl) { 8773 unsigned OpNum = (PFEntry >> 26) & 0x0F; 8774 unsigned LHSID = (PFEntry >> 13) & ((1 << 13) - 1); 8775 unsigned RHSID = (PFEntry >> 0) & ((1 << 13) - 1); 8776 8777 enum { 8778 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3> 8779 OP_VREV, 8780 OP_VDUP0, 8781 OP_VDUP1, 8782 OP_VDUP2, 8783 OP_VDUP3, 8784 OP_VEXT1, 8785 OP_VEXT2, 8786 OP_VEXT3, 8787 OP_VUZPL, // VUZP, left result 8788 OP_VUZPR, // VUZP, right result 8789 OP_VZIPL, // VZIP, left result 8790 OP_VZIPR, // VZIP, right result 8791 OP_VTRNL, // VTRN, left result 8792 OP_VTRNR // VTRN, right result 8793 }; 8794 8795 if (OpNum == OP_COPY) { 8796 if (LHSID == (1 * 9 + 2) * 9 + 3) 8797 return LHS; 8798 assert(LHSID == ((4 * 9 + 5) * 9 + 6) * 9 + 7 && "Illegal OP_COPY!"); 8799 return RHS; 8800 } 8801 8802 SDValue OpLHS, OpRHS; 8803 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl); 8804 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl); 8805 EVT VT = OpLHS.getValueType(); 8806 8807 switch (OpNum) { 8808 default: 8809 llvm_unreachable("Unknown shuffle opcode!"); 8810 case OP_VREV: 8811 // VREV divides the vector in half and swaps within the half. 8812 if (VT.getVectorElementType() == MVT::i32 || 8813 VT.getVectorElementType() == MVT::f32) 8814 return DAG.getNode(AArch64ISD::REV64, dl, VT, OpLHS); 8815 // vrev <4 x i16> -> REV32 8816 if (VT.getVectorElementType() == MVT::i16 || 8817 VT.getVectorElementType() == MVT::f16 || 8818 VT.getVectorElementType() == MVT::bf16) 8819 return DAG.getNode(AArch64ISD::REV32, dl, VT, OpLHS); 8820 // vrev <4 x i8> -> REV16 8821 assert(VT.getVectorElementType() == MVT::i8); 8822 return DAG.getNode(AArch64ISD::REV16, dl, VT, OpLHS); 8823 case OP_VDUP0: 8824 case OP_VDUP1: 8825 case OP_VDUP2: 8826 case OP_VDUP3: { 8827 EVT EltTy = VT.getVectorElementType(); 8828 unsigned Opcode; 8829 if (EltTy == MVT::i8) 8830 Opcode = AArch64ISD::DUPLANE8; 8831 else if (EltTy == MVT::i16 || EltTy == MVT::f16 || EltTy == MVT::bf16) 8832 Opcode = AArch64ISD::DUPLANE16; 8833 else if (EltTy == MVT::i32 || EltTy == MVT::f32) 8834 Opcode = AArch64ISD::DUPLANE32; 8835 else if (EltTy == MVT::i64 || EltTy == MVT::f64) 8836 Opcode = AArch64ISD::DUPLANE64; 8837 else 8838 llvm_unreachable("Invalid vector element type?"); 8839 8840 if (VT.getSizeInBits() == 64) 8841 OpLHS = WidenVector(OpLHS, DAG); 8842 SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, dl, MVT::i64); 8843 return DAG.getNode(Opcode, dl, VT, OpLHS, Lane); 8844 } 8845 case OP_VEXT1: 8846 case OP_VEXT2: 8847 case OP_VEXT3: { 8848 unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS); 8849 return DAG.getNode(AArch64ISD::EXT, dl, VT, OpLHS, OpRHS, 8850 DAG.getConstant(Imm, dl, MVT::i32)); 8851 } 8852 case OP_VUZPL: 8853 return DAG.getNode(AArch64ISD::UZP1, dl, DAG.getVTList(VT, VT), OpLHS, 8854 OpRHS); 8855 case OP_VUZPR: 8856 return DAG.getNode(AArch64ISD::UZP2, dl, DAG.getVTList(VT, VT), OpLHS, 8857 OpRHS); 8858 case OP_VZIPL: 8859 return DAG.getNode(AArch64ISD::ZIP1, dl, DAG.getVTList(VT, VT), OpLHS, 8860 OpRHS); 8861 case OP_VZIPR: 8862 return DAG.getNode(AArch64ISD::ZIP2, dl, DAG.getVTList(VT, VT), OpLHS, 8863 OpRHS); 8864 case OP_VTRNL: 8865 return DAG.getNode(AArch64ISD::TRN1, dl, DAG.getVTList(VT, VT), OpLHS, 8866 OpRHS); 8867 case OP_VTRNR: 8868 return DAG.getNode(AArch64ISD::TRN2, dl, DAG.getVTList(VT, VT), OpLHS, 8869 OpRHS); 8870 } 8871 } 8872 8873 static SDValue GenerateTBL(SDValue Op, ArrayRef<int> ShuffleMask, 8874 SelectionDAG &DAG) { 8875 // Check to see if we can use the TBL instruction. 8876 SDValue V1 = Op.getOperand(0); 8877 SDValue V2 = Op.getOperand(1); 8878 SDLoc DL(Op); 8879 8880 EVT EltVT = Op.getValueType().getVectorElementType(); 8881 unsigned BytesPerElt = EltVT.getSizeInBits() / 8; 8882 8883 SmallVector<SDValue, 8> TBLMask; 8884 for (int Val : ShuffleMask) { 8885 for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) { 8886 unsigned Offset = Byte + Val * BytesPerElt; 8887 TBLMask.push_back(DAG.getConstant(Offset, DL, MVT::i32)); 8888 } 8889 } 8890 8891 MVT IndexVT = MVT::v8i8; 8892 unsigned IndexLen = 8; 8893 if (Op.getValueSizeInBits() == 128) { 8894 IndexVT = MVT::v16i8; 8895 IndexLen = 16; 8896 } 8897 8898 SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1); 8899 SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2); 8900 8901 SDValue Shuffle; 8902 if (V2.getNode()->isUndef()) { 8903 if (IndexLen == 8) 8904 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst); 8905 Shuffle = DAG.getNode( 8906 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT, 8907 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, DL, MVT::i32), V1Cst, 8908 DAG.getBuildVector(IndexVT, DL, 8909 makeArrayRef(TBLMask.data(), IndexLen))); 8910 } else { 8911 if (IndexLen == 8) { 8912 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst); 8913 Shuffle = DAG.getNode( 8914 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT, 8915 DAG.getConstant(Intrinsic::aarch64_neon_tbl1, DL, MVT::i32), V1Cst, 8916 DAG.getBuildVector(IndexVT, DL, 8917 makeArrayRef(TBLMask.data(), IndexLen))); 8918 } else { 8919 // FIXME: We cannot, for the moment, emit a TBL2 instruction because we 8920 // cannot currently represent the register constraints on the input 8921 // table registers. 8922 // Shuffle = DAG.getNode(AArch64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst, 8923 // DAG.getBuildVector(IndexVT, DL, &TBLMask[0], 8924 // IndexLen)); 8925 Shuffle = DAG.getNode( 8926 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT, 8927 DAG.getConstant(Intrinsic::aarch64_neon_tbl2, DL, MVT::i32), V1Cst, 8928 V2Cst, DAG.getBuildVector(IndexVT, DL, 8929 makeArrayRef(TBLMask.data(), IndexLen))); 8930 } 8931 } 8932 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle); 8933 } 8934 8935 static unsigned getDUPLANEOp(EVT EltType) { 8936 if (EltType == MVT::i8) 8937 return AArch64ISD::DUPLANE8; 8938 if (EltType == MVT::i16 || EltType == MVT::f16 || EltType == MVT::bf16) 8939 return AArch64ISD::DUPLANE16; 8940 if (EltType == MVT::i32 || EltType == MVT::f32) 8941 return AArch64ISD::DUPLANE32; 8942 if (EltType == MVT::i64 || EltType == MVT::f64) 8943 return AArch64ISD::DUPLANE64; 8944 8945 llvm_unreachable("Invalid vector element type?"); 8946 } 8947 8948 static SDValue constructDup(SDValue V, int Lane, SDLoc dl, EVT VT, 8949 unsigned Opcode, SelectionDAG &DAG) { 8950 // Try to eliminate a bitcasted extract subvector before a DUPLANE. 8951 auto getScaledOffsetDup = [](SDValue BitCast, int &LaneC, MVT &CastVT) { 8952 // Match: dup (bitcast (extract_subv X, C)), LaneC 8953 if (BitCast.getOpcode() != ISD::BITCAST || 8954 BitCast.getOperand(0).getOpcode() != ISD::EXTRACT_SUBVECTOR) 8955 return false; 8956 8957 // The extract index must align in the destination type. That may not 8958 // happen if the bitcast is from narrow to wide type. 8959 SDValue Extract = BitCast.getOperand(0); 8960 unsigned ExtIdx = Extract.getConstantOperandVal(1); 8961 unsigned SrcEltBitWidth = Extract.getScalarValueSizeInBits(); 8962 unsigned ExtIdxInBits = ExtIdx * SrcEltBitWidth; 8963 unsigned CastedEltBitWidth = BitCast.getScalarValueSizeInBits(); 8964 if (ExtIdxInBits % CastedEltBitWidth != 0) 8965 return false; 8966 8967 // Update the lane value by offsetting with the scaled extract index. 8968 LaneC += ExtIdxInBits / CastedEltBitWidth; 8969 8970 // Determine the casted vector type of the wide vector input. 8971 // dup (bitcast (extract_subv X, C)), LaneC --> dup (bitcast X), LaneC' 8972 // Examples: 8973 // dup (bitcast (extract_subv v2f64 X, 1) to v2f32), 1 --> dup v4f32 X, 3 8974 // dup (bitcast (extract_subv v16i8 X, 8) to v4i16), 1 --> dup v8i16 X, 5 8975 unsigned SrcVecNumElts = 8976 Extract.getOperand(0).getValueSizeInBits() / CastedEltBitWidth; 8977 CastVT = MVT::getVectorVT(BitCast.getSimpleValueType().getScalarType(), 8978 SrcVecNumElts); 8979 return true; 8980 }; 8981 MVT CastVT; 8982 if (getScaledOffsetDup(V, Lane, CastVT)) { 8983 V = DAG.getBitcast(CastVT, V.getOperand(0).getOperand(0)); 8984 } else if (V.getOpcode() == ISD::EXTRACT_SUBVECTOR) { 8985 // The lane is incremented by the index of the extract. 8986 // Example: dup v2f32 (extract v4f32 X, 2), 1 --> dup v4f32 X, 3 8987 Lane += V.getConstantOperandVal(1); 8988 V = V.getOperand(0); 8989 } else if (V.getOpcode() == ISD::CONCAT_VECTORS) { 8990 // The lane is decremented if we are splatting from the 2nd operand. 8991 // Example: dup v4i32 (concat v2i32 X, v2i32 Y), 3 --> dup v4i32 Y, 1 8992 unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2; 8993 Lane -= Idx * VT.getVectorNumElements() / 2; 8994 V = WidenVector(V.getOperand(Idx), DAG); 8995 } else if (VT.getSizeInBits() == 64) { 8996 // Widen the operand to 128-bit register with undef. 8997 V = WidenVector(V, DAG); 8998 } 8999 return DAG.getNode(Opcode, dl, VT, V, DAG.getConstant(Lane, dl, MVT::i64)); 9000 } 9001 9002 SDValue AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, 9003 SelectionDAG &DAG) const { 9004 SDLoc dl(Op); 9005 EVT VT = Op.getValueType(); 9006 9007 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode()); 9008 9009 // Convert shuffles that are directly supported on NEON to target-specific 9010 // DAG nodes, instead of keeping them as shuffles and matching them again 9011 // during code selection. This is more efficient and avoids the possibility 9012 // of inconsistencies between legalization and selection. 9013 ArrayRef<int> ShuffleMask = SVN->getMask(); 9014 9015 SDValue V1 = Op.getOperand(0); 9016 SDValue V2 = Op.getOperand(1); 9017 9018 assert(V1.getValueType() == VT && "Unexpected VECTOR_SHUFFLE type!"); 9019 assert(ShuffleMask.size() == VT.getVectorNumElements() && 9020 "Unexpected VECTOR_SHUFFLE mask size!"); 9021 9022 if (SVN->isSplat()) { 9023 int Lane = SVN->getSplatIndex(); 9024 // If this is undef splat, generate it via "just" vdup, if possible. 9025 if (Lane == -1) 9026 Lane = 0; 9027 9028 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) 9029 return DAG.getNode(AArch64ISD::DUP, dl, V1.getValueType(), 9030 V1.getOperand(0)); 9031 // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non- 9032 // constant. If so, we can just reference the lane's definition directly. 9033 if (V1.getOpcode() == ISD::BUILD_VECTOR && 9034 !isa<ConstantSDNode>(V1.getOperand(Lane))) 9035 return DAG.getNode(AArch64ISD::DUP, dl, VT, V1.getOperand(Lane)); 9036 9037 // Otherwise, duplicate from the lane of the input vector. 9038 unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType()); 9039 return constructDup(V1, Lane, dl, VT, Opcode, DAG); 9040 } 9041 9042 // Check if the mask matches a DUP for a wider element 9043 for (unsigned LaneSize : {64U, 32U, 16U}) { 9044 unsigned Lane = 0; 9045 if (isWideDUPMask(ShuffleMask, VT, LaneSize, Lane)) { 9046 unsigned Opcode = LaneSize == 64 ? AArch64ISD::DUPLANE64 9047 : LaneSize == 32 ? AArch64ISD::DUPLANE32 9048 : AArch64ISD::DUPLANE16; 9049 // Cast V1 to an integer vector with required lane size 9050 MVT NewEltTy = MVT::getIntegerVT(LaneSize); 9051 unsigned NewEltCount = VT.getSizeInBits() / LaneSize; 9052 MVT NewVecTy = MVT::getVectorVT(NewEltTy, NewEltCount); 9053 V1 = DAG.getBitcast(NewVecTy, V1); 9054 // Constuct the DUP instruction 9055 V1 = constructDup(V1, Lane, dl, NewVecTy, Opcode, DAG); 9056 // Cast back to the original type 9057 return DAG.getBitcast(VT, V1); 9058 } 9059 } 9060 9061 if (isREVMask(ShuffleMask, VT, 64)) 9062 return DAG.getNode(AArch64ISD::REV64, dl, V1.getValueType(), V1, V2); 9063 if (isREVMask(ShuffleMask, VT, 32)) 9064 return DAG.getNode(AArch64ISD::REV32, dl, V1.getValueType(), V1, V2); 9065 if (isREVMask(ShuffleMask, VT, 16)) 9066 return DAG.getNode(AArch64ISD::REV16, dl, V1.getValueType(), V1, V2); 9067 9068 if (((VT.getVectorNumElements() == 8 && VT.getScalarSizeInBits() == 16) || 9069 (VT.getVectorNumElements() == 16 && VT.getScalarSizeInBits() == 8)) && 9070 ShuffleVectorInst::isReverseMask(ShuffleMask)) { 9071 SDValue Rev = DAG.getNode(AArch64ISD::REV64, dl, VT, V1); 9072 return DAG.getNode(AArch64ISD::EXT, dl, VT, Rev, Rev, 9073 DAG.getConstant(8, dl, MVT::i32)); 9074 } 9075 9076 bool ReverseEXT = false; 9077 unsigned Imm; 9078 if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) { 9079 if (ReverseEXT) 9080 std::swap(V1, V2); 9081 Imm *= getExtFactor(V1); 9082 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V2, 9083 DAG.getConstant(Imm, dl, MVT::i32)); 9084 } else if (V2->isUndef() && isSingletonEXTMask(ShuffleMask, VT, Imm)) { 9085 Imm *= getExtFactor(V1); 9086 return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V1, 9087 DAG.getConstant(Imm, dl, MVT::i32)); 9088 } 9089 9090 unsigned WhichResult; 9091 if (isZIPMask(ShuffleMask, VT, WhichResult)) { 9092 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2; 9093 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2); 9094 } 9095 if (isUZPMask(ShuffleMask, VT, WhichResult)) { 9096 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2; 9097 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2); 9098 } 9099 if (isTRNMask(ShuffleMask, VT, WhichResult)) { 9100 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2; 9101 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2); 9102 } 9103 9104 if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) { 9105 unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2; 9106 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1); 9107 } 9108 if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) { 9109 unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2; 9110 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1); 9111 } 9112 if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) { 9113 unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2; 9114 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1); 9115 } 9116 9117 if (SDValue Concat = tryFormConcatFromShuffle(Op, DAG)) 9118 return Concat; 9119 9120 bool DstIsLeft; 9121 int Anomaly; 9122 int NumInputElements = V1.getValueType().getVectorNumElements(); 9123 if (isINSMask(ShuffleMask, NumInputElements, DstIsLeft, Anomaly)) { 9124 SDValue DstVec = DstIsLeft ? V1 : V2; 9125 SDValue DstLaneV = DAG.getConstant(Anomaly, dl, MVT::i64); 9126 9127 SDValue SrcVec = V1; 9128 int SrcLane = ShuffleMask[Anomaly]; 9129 if (SrcLane >= NumInputElements) { 9130 SrcVec = V2; 9131 SrcLane -= VT.getVectorNumElements(); 9132 } 9133 SDValue SrcLaneV = DAG.getConstant(SrcLane, dl, MVT::i64); 9134 9135 EVT ScalarVT = VT.getVectorElementType(); 9136 9137 if (ScalarVT.getFixedSizeInBits() < 32 && ScalarVT.isInteger()) 9138 ScalarVT = MVT::i32; 9139 9140 return DAG.getNode( 9141 ISD::INSERT_VECTOR_ELT, dl, VT, DstVec, 9142 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ScalarVT, SrcVec, SrcLaneV), 9143 DstLaneV); 9144 } 9145 9146 // If the shuffle is not directly supported and it has 4 elements, use 9147 // the PerfectShuffle-generated table to synthesize it from other shuffles. 9148 unsigned NumElts = VT.getVectorNumElements(); 9149 if (NumElts == 4) { 9150 unsigned PFIndexes[4]; 9151 for (unsigned i = 0; i != 4; ++i) { 9152 if (ShuffleMask[i] < 0) 9153 PFIndexes[i] = 8; 9154 else 9155 PFIndexes[i] = ShuffleMask[i]; 9156 } 9157 9158 // Compute the index in the perfect shuffle table. 9159 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 + 9160 PFIndexes[2] * 9 + PFIndexes[3]; 9161 unsigned PFEntry = PerfectShuffleTable[PFTableIndex]; 9162 unsigned Cost = (PFEntry >> 30); 9163 9164 if (Cost <= 4) 9165 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl); 9166 } 9167 9168 return GenerateTBL(Op, ShuffleMask, DAG); 9169 } 9170 9171 SDValue AArch64TargetLowering::LowerSPLAT_VECTOR(SDValue Op, 9172 SelectionDAG &DAG) const { 9173 SDLoc dl(Op); 9174 EVT VT = Op.getValueType(); 9175 EVT ElemVT = VT.getScalarType(); 9176 SDValue SplatVal = Op.getOperand(0); 9177 9178 if (useSVEForFixedLengthVectorVT(VT)) 9179 return LowerToScalableOp(Op, DAG); 9180 9181 // Extend input splat value where needed to fit into a GPR (32b or 64b only) 9182 // FPRs don't have this restriction. 9183 switch (ElemVT.getSimpleVT().SimpleTy) { 9184 case MVT::i1: { 9185 // The only legal i1 vectors are SVE vectors, so we can use SVE-specific 9186 // lowering code. 9187 if (auto *ConstVal = dyn_cast<ConstantSDNode>(SplatVal)) { 9188 if (ConstVal->isOne()) 9189 return getPTrue(DAG, dl, VT, AArch64SVEPredPattern::all); 9190 // TODO: Add special case for constant false 9191 } 9192 // The general case of i1. There isn't any natural way to do this, 9193 // so we use some trickery with whilelo. 9194 SplatVal = DAG.getAnyExtOrTrunc(SplatVal, dl, MVT::i64); 9195 SplatVal = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::i64, SplatVal, 9196 DAG.getValueType(MVT::i1)); 9197 SDValue ID = DAG.getTargetConstant(Intrinsic::aarch64_sve_whilelo, dl, 9198 MVT::i64); 9199 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT, ID, 9200 DAG.getConstant(0, dl, MVT::i64), SplatVal); 9201 } 9202 case MVT::i8: 9203 case MVT::i16: 9204 case MVT::i32: 9205 SplatVal = DAG.getAnyExtOrTrunc(SplatVal, dl, MVT::i32); 9206 break; 9207 case MVT::i64: 9208 SplatVal = DAG.getAnyExtOrTrunc(SplatVal, dl, MVT::i64); 9209 break; 9210 case MVT::f16: 9211 case MVT::bf16: 9212 case MVT::f32: 9213 case MVT::f64: 9214 // Fine as is 9215 break; 9216 default: 9217 report_fatal_error("Unsupported SPLAT_VECTOR input operand type"); 9218 } 9219 9220 return DAG.getNode(AArch64ISD::DUP, dl, VT, SplatVal); 9221 } 9222 9223 SDValue AArch64TargetLowering::LowerDUPQLane(SDValue Op, 9224 SelectionDAG &DAG) const { 9225 SDLoc DL(Op); 9226 9227 EVT VT = Op.getValueType(); 9228 if (!isTypeLegal(VT) || !VT.isScalableVector()) 9229 return SDValue(); 9230 9231 // Current lowering only supports the SVE-ACLE types. 9232 if (VT.getSizeInBits().getKnownMinSize() != AArch64::SVEBitsPerBlock) 9233 return SDValue(); 9234 9235 // The DUPQ operation is indepedent of element type so normalise to i64s. 9236 SDValue V = DAG.getNode(ISD::BITCAST, DL, MVT::nxv2i64, Op.getOperand(1)); 9237 SDValue Idx128 = Op.getOperand(2); 9238 9239 // DUPQ can be used when idx is in range. 9240 auto *CIdx = dyn_cast<ConstantSDNode>(Idx128); 9241 if (CIdx && (CIdx->getZExtValue() <= 3)) { 9242 SDValue CI = DAG.getTargetConstant(CIdx->getZExtValue(), DL, MVT::i64); 9243 SDNode *DUPQ = 9244 DAG.getMachineNode(AArch64::DUP_ZZI_Q, DL, MVT::nxv2i64, V, CI); 9245 return DAG.getNode(ISD::BITCAST, DL, VT, SDValue(DUPQ, 0)); 9246 } 9247 9248 // The ACLE says this must produce the same result as: 9249 // svtbl(data, svadd_x(svptrue_b64(), 9250 // svand_x(svptrue_b64(), svindex_u64(0, 1), 1), 9251 // index * 2)) 9252 SDValue One = DAG.getConstant(1, DL, MVT::i64); 9253 SDValue SplatOne = DAG.getNode(ISD::SPLAT_VECTOR, DL, MVT::nxv2i64, One); 9254 9255 // create the vector 0,1,0,1,... 9256 SDValue SV = DAG.getNode(ISD::STEP_VECTOR, DL, MVT::nxv2i64, One); 9257 SV = DAG.getNode(ISD::AND, DL, MVT::nxv2i64, SV, SplatOne); 9258 9259 // create the vector idx64,idx64+1,idx64,idx64+1,... 9260 SDValue Idx64 = DAG.getNode(ISD::ADD, DL, MVT::i64, Idx128, Idx128); 9261 SDValue SplatIdx64 = DAG.getNode(ISD::SPLAT_VECTOR, DL, MVT::nxv2i64, Idx64); 9262 SDValue ShuffleMask = DAG.getNode(ISD::ADD, DL, MVT::nxv2i64, SV, SplatIdx64); 9263 9264 // create the vector Val[idx64],Val[idx64+1],Val[idx64],Val[idx64+1],... 9265 SDValue TBL = DAG.getNode(AArch64ISD::TBL, DL, MVT::nxv2i64, V, ShuffleMask); 9266 return DAG.getNode(ISD::BITCAST, DL, VT, TBL); 9267 } 9268 9269 9270 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits, 9271 APInt &UndefBits) { 9272 EVT VT = BVN->getValueType(0); 9273 APInt SplatBits, SplatUndef; 9274 unsigned SplatBitSize; 9275 bool HasAnyUndefs; 9276 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) { 9277 unsigned NumSplats = VT.getSizeInBits() / SplatBitSize; 9278 9279 for (unsigned i = 0; i < NumSplats; ++i) { 9280 CnstBits <<= SplatBitSize; 9281 UndefBits <<= SplatBitSize; 9282 CnstBits |= SplatBits.zextOrTrunc(VT.getSizeInBits()); 9283 UndefBits |= (SplatBits ^ SplatUndef).zextOrTrunc(VT.getSizeInBits()); 9284 } 9285 9286 return true; 9287 } 9288 9289 return false; 9290 } 9291 9292 // Try 64-bit splatted SIMD immediate. 9293 static SDValue tryAdvSIMDModImm64(unsigned NewOp, SDValue Op, SelectionDAG &DAG, 9294 const APInt &Bits) { 9295 if (Bits.getHiBits(64) == Bits.getLoBits(64)) { 9296 uint64_t Value = Bits.zextOrTrunc(64).getZExtValue(); 9297 EVT VT = Op.getValueType(); 9298 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v2i64 : MVT::f64; 9299 9300 if (AArch64_AM::isAdvSIMDModImmType10(Value)) { 9301 Value = AArch64_AM::encodeAdvSIMDModImmType10(Value); 9302 9303 SDLoc dl(Op); 9304 SDValue Mov = DAG.getNode(NewOp, dl, MovTy, 9305 DAG.getConstant(Value, dl, MVT::i32)); 9306 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov); 9307 } 9308 } 9309 9310 return SDValue(); 9311 } 9312 9313 // Try 32-bit splatted SIMD immediate. 9314 static SDValue tryAdvSIMDModImm32(unsigned NewOp, SDValue Op, SelectionDAG &DAG, 9315 const APInt &Bits, 9316 const SDValue *LHS = nullptr) { 9317 if (Bits.getHiBits(64) == Bits.getLoBits(64)) { 9318 uint64_t Value = Bits.zextOrTrunc(64).getZExtValue(); 9319 EVT VT = Op.getValueType(); 9320 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32; 9321 bool isAdvSIMDModImm = false; 9322 uint64_t Shift; 9323 9324 if ((isAdvSIMDModImm = AArch64_AM::isAdvSIMDModImmType1(Value))) { 9325 Value = AArch64_AM::encodeAdvSIMDModImmType1(Value); 9326 Shift = 0; 9327 } 9328 else if ((isAdvSIMDModImm = AArch64_AM::isAdvSIMDModImmType2(Value))) { 9329 Value = AArch64_AM::encodeAdvSIMDModImmType2(Value); 9330 Shift = 8; 9331 } 9332 else if ((isAdvSIMDModImm = AArch64_AM::isAdvSIMDModImmType3(Value))) { 9333 Value = AArch64_AM::encodeAdvSIMDModImmType3(Value); 9334 Shift = 16; 9335 } 9336 else if ((isAdvSIMDModImm = AArch64_AM::isAdvSIMDModImmType4(Value))) { 9337 Value = AArch64_AM::encodeAdvSIMDModImmType4(Value); 9338 Shift = 24; 9339 } 9340 9341 if (isAdvSIMDModImm) { 9342 SDLoc dl(Op); 9343 SDValue Mov; 9344 9345 if (LHS) 9346 Mov = DAG.getNode(NewOp, dl, MovTy, *LHS, 9347 DAG.getConstant(Value, dl, MVT::i32), 9348 DAG.getConstant(Shift, dl, MVT::i32)); 9349 else 9350 Mov = DAG.getNode(NewOp, dl, MovTy, 9351 DAG.getConstant(Value, dl, MVT::i32), 9352 DAG.getConstant(Shift, dl, MVT::i32)); 9353 9354 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov); 9355 } 9356 } 9357 9358 return SDValue(); 9359 } 9360 9361 // Try 16-bit splatted SIMD immediate. 9362 static SDValue tryAdvSIMDModImm16(unsigned NewOp, SDValue Op, SelectionDAG &DAG, 9363 const APInt &Bits, 9364 const SDValue *LHS = nullptr) { 9365 if (Bits.getHiBits(64) == Bits.getLoBits(64)) { 9366 uint64_t Value = Bits.zextOrTrunc(64).getZExtValue(); 9367 EVT VT = Op.getValueType(); 9368 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16; 9369 bool isAdvSIMDModImm = false; 9370 uint64_t Shift; 9371 9372 if ((isAdvSIMDModImm = AArch64_AM::isAdvSIMDModImmType5(Value))) { 9373 Value = AArch64_AM::encodeAdvSIMDModImmType5(Value); 9374 Shift = 0; 9375 } 9376 else if ((isAdvSIMDModImm = AArch64_AM::isAdvSIMDModImmType6(Value))) { 9377 Value = AArch64_AM::encodeAdvSIMDModImmType6(Value); 9378 Shift = 8; 9379 } 9380 9381 if (isAdvSIMDModImm) { 9382 SDLoc dl(Op); 9383 SDValue Mov; 9384 9385 if (LHS) 9386 Mov = DAG.getNode(NewOp, dl, MovTy, *LHS, 9387 DAG.getConstant(Value, dl, MVT::i32), 9388 DAG.getConstant(Shift, dl, MVT::i32)); 9389 else 9390 Mov = DAG.getNode(NewOp, dl, MovTy, 9391 DAG.getConstant(Value, dl, MVT::i32), 9392 DAG.getConstant(Shift, dl, MVT::i32)); 9393 9394 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov); 9395 } 9396 } 9397 9398 return SDValue(); 9399 } 9400 9401 // Try 32-bit splatted SIMD immediate with shifted ones. 9402 static SDValue tryAdvSIMDModImm321s(unsigned NewOp, SDValue Op, 9403 SelectionDAG &DAG, const APInt &Bits) { 9404 if (Bits.getHiBits(64) == Bits.getLoBits(64)) { 9405 uint64_t Value = Bits.zextOrTrunc(64).getZExtValue(); 9406 EVT VT = Op.getValueType(); 9407 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32; 9408 bool isAdvSIMDModImm = false; 9409 uint64_t Shift; 9410 9411 if ((isAdvSIMDModImm = AArch64_AM::isAdvSIMDModImmType7(Value))) { 9412 Value = AArch64_AM::encodeAdvSIMDModImmType7(Value); 9413 Shift = 264; 9414 } 9415 else if ((isAdvSIMDModImm = AArch64_AM::isAdvSIMDModImmType8(Value))) { 9416 Value = AArch64_AM::encodeAdvSIMDModImmType8(Value); 9417 Shift = 272; 9418 } 9419 9420 if (isAdvSIMDModImm) { 9421 SDLoc dl(Op); 9422 SDValue Mov = DAG.getNode(NewOp, dl, MovTy, 9423 DAG.getConstant(Value, dl, MVT::i32), 9424 DAG.getConstant(Shift, dl, MVT::i32)); 9425 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov); 9426 } 9427 } 9428 9429 return SDValue(); 9430 } 9431 9432 // Try 8-bit splatted SIMD immediate. 9433 static SDValue tryAdvSIMDModImm8(unsigned NewOp, SDValue Op, SelectionDAG &DAG, 9434 const APInt &Bits) { 9435 if (Bits.getHiBits(64) == Bits.getLoBits(64)) { 9436 uint64_t Value = Bits.zextOrTrunc(64).getZExtValue(); 9437 EVT VT = Op.getValueType(); 9438 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8; 9439 9440 if (AArch64_AM::isAdvSIMDModImmType9(Value)) { 9441 Value = AArch64_AM::encodeAdvSIMDModImmType9(Value); 9442 9443 SDLoc dl(Op); 9444 SDValue Mov = DAG.getNode(NewOp, dl, MovTy, 9445 DAG.getConstant(Value, dl, MVT::i32)); 9446 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov); 9447 } 9448 } 9449 9450 return SDValue(); 9451 } 9452 9453 // Try FP splatted SIMD immediate. 9454 static SDValue tryAdvSIMDModImmFP(unsigned NewOp, SDValue Op, SelectionDAG &DAG, 9455 const APInt &Bits) { 9456 if (Bits.getHiBits(64) == Bits.getLoBits(64)) { 9457 uint64_t Value = Bits.zextOrTrunc(64).getZExtValue(); 9458 EVT VT = Op.getValueType(); 9459 bool isWide = (VT.getSizeInBits() == 128); 9460 MVT MovTy; 9461 bool isAdvSIMDModImm = false; 9462 9463 if ((isAdvSIMDModImm = AArch64_AM::isAdvSIMDModImmType11(Value))) { 9464 Value = AArch64_AM::encodeAdvSIMDModImmType11(Value); 9465 MovTy = isWide ? MVT::v4f32 : MVT::v2f32; 9466 } 9467 else if (isWide && 9468 (isAdvSIMDModImm = AArch64_AM::isAdvSIMDModImmType12(Value))) { 9469 Value = AArch64_AM::encodeAdvSIMDModImmType12(Value); 9470 MovTy = MVT::v2f64; 9471 } 9472 9473 if (isAdvSIMDModImm) { 9474 SDLoc dl(Op); 9475 SDValue Mov = DAG.getNode(NewOp, dl, MovTy, 9476 DAG.getConstant(Value, dl, MVT::i32)); 9477 return DAG.getNode(AArch64ISD::NVCAST, dl, VT, Mov); 9478 } 9479 } 9480 9481 return SDValue(); 9482 } 9483 9484 // Specialized code to quickly find if PotentialBVec is a BuildVector that 9485 // consists of only the same constant int value, returned in reference arg 9486 // ConstVal 9487 static bool isAllConstantBuildVector(const SDValue &PotentialBVec, 9488 uint64_t &ConstVal) { 9489 BuildVectorSDNode *Bvec = dyn_cast<BuildVectorSDNode>(PotentialBVec); 9490 if (!Bvec) 9491 return false; 9492 ConstantSDNode *FirstElt = dyn_cast<ConstantSDNode>(Bvec->getOperand(0)); 9493 if (!FirstElt) 9494 return false; 9495 EVT VT = Bvec->getValueType(0); 9496 unsigned NumElts = VT.getVectorNumElements(); 9497 for (unsigned i = 1; i < NumElts; ++i) 9498 if (dyn_cast<ConstantSDNode>(Bvec->getOperand(i)) != FirstElt) 9499 return false; 9500 ConstVal = FirstElt->getZExtValue(); 9501 return true; 9502 } 9503 9504 static unsigned getIntrinsicID(const SDNode *N) { 9505 unsigned Opcode = N->getOpcode(); 9506 switch (Opcode) { 9507 default: 9508 return Intrinsic::not_intrinsic; 9509 case ISD::INTRINSIC_WO_CHAIN: { 9510 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue(); 9511 if (IID < Intrinsic::num_intrinsics) 9512 return IID; 9513 return Intrinsic::not_intrinsic; 9514 } 9515 } 9516 } 9517 9518 // Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)), 9519 // to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a 9520 // BUILD_VECTORs with constant element C1, C2 is a constant, and: 9521 // - for the SLI case: C1 == ~(Ones(ElemSizeInBits) << C2) 9522 // - for the SRI case: C1 == ~(Ones(ElemSizeInBits) >> C2) 9523 // The (or (lsl Y, C2), (and X, BvecC1)) case is also handled. 9524 static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) { 9525 EVT VT = N->getValueType(0); 9526 9527 if (!VT.isVector()) 9528 return SDValue(); 9529 9530 SDLoc DL(N); 9531 9532 SDValue And; 9533 SDValue Shift; 9534 9535 SDValue FirstOp = N->getOperand(0); 9536 unsigned FirstOpc = FirstOp.getOpcode(); 9537 SDValue SecondOp = N->getOperand(1); 9538 unsigned SecondOpc = SecondOp.getOpcode(); 9539 9540 // Is one of the operands an AND or a BICi? The AND may have been optimised to 9541 // a BICi in order to use an immediate instead of a register. 9542 // Is the other operand an shl or lshr? This will have been turned into: 9543 // AArch64ISD::VSHL vector, #shift or AArch64ISD::VLSHR vector, #shift. 9544 if ((FirstOpc == ISD::AND || FirstOpc == AArch64ISD::BICi) && 9545 (SecondOpc == AArch64ISD::VSHL || SecondOpc == AArch64ISD::VLSHR)) { 9546 And = FirstOp; 9547 Shift = SecondOp; 9548 9549 } else if ((SecondOpc == ISD::AND || SecondOpc == AArch64ISD::BICi) && 9550 (FirstOpc == AArch64ISD::VSHL || FirstOpc == AArch64ISD::VLSHR)) { 9551 And = SecondOp; 9552 Shift = FirstOp; 9553 } else 9554 return SDValue(); 9555 9556 bool IsAnd = And.getOpcode() == ISD::AND; 9557 bool IsShiftRight = Shift.getOpcode() == AArch64ISD::VLSHR; 9558 9559 // Is the shift amount constant? 9560 ConstantSDNode *C2node = dyn_cast<ConstantSDNode>(Shift.getOperand(1)); 9561 if (!C2node) 9562 return SDValue(); 9563 9564 uint64_t C1; 9565 if (IsAnd) { 9566 // Is the and mask vector all constant? 9567 if (!isAllConstantBuildVector(And.getOperand(1), C1)) 9568 return SDValue(); 9569 } else { 9570 // Reconstruct the corresponding AND immediate from the two BICi immediates. 9571 ConstantSDNode *C1nodeImm = dyn_cast<ConstantSDNode>(And.getOperand(1)); 9572 ConstantSDNode *C1nodeShift = dyn_cast<ConstantSDNode>(And.getOperand(2)); 9573 assert(C1nodeImm && C1nodeShift); 9574 C1 = ~(C1nodeImm->getZExtValue() << C1nodeShift->getZExtValue()); 9575 } 9576 9577 // Is C1 == ~(Ones(ElemSizeInBits) << C2) or 9578 // C1 == ~(Ones(ElemSizeInBits) >> C2), taking into account 9579 // how much one can shift elements of a particular size? 9580 uint64_t C2 = C2node->getZExtValue(); 9581 unsigned ElemSizeInBits = VT.getScalarSizeInBits(); 9582 if (C2 > ElemSizeInBits) 9583 return SDValue(); 9584 9585 APInt C1AsAPInt(ElemSizeInBits, C1); 9586 APInt RequiredC1 = IsShiftRight ? APInt::getHighBitsSet(ElemSizeInBits, C2) 9587 : APInt::getLowBitsSet(ElemSizeInBits, C2); 9588 if (C1AsAPInt != RequiredC1) 9589 return SDValue(); 9590 9591 SDValue X = And.getOperand(0); 9592 SDValue Y = Shift.getOperand(0); 9593 9594 unsigned Inst = IsShiftRight ? AArch64ISD::VSRI : AArch64ISD::VSLI; 9595 SDValue ResultSLI = DAG.getNode(Inst, DL, VT, X, Y, Shift.getOperand(1)); 9596 9597 LLVM_DEBUG(dbgs() << "aarch64-lower: transformed: \n"); 9598 LLVM_DEBUG(N->dump(&DAG)); 9599 LLVM_DEBUG(dbgs() << "into: \n"); 9600 LLVM_DEBUG(ResultSLI->dump(&DAG)); 9601 9602 ++NumShiftInserts; 9603 return ResultSLI; 9604 } 9605 9606 SDValue AArch64TargetLowering::LowerVectorOR(SDValue Op, 9607 SelectionDAG &DAG) const { 9608 if (useSVEForFixedLengthVectorVT(Op.getValueType())) 9609 return LowerToScalableOp(Op, DAG); 9610 9611 // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2)) 9612 if (SDValue Res = tryLowerToSLI(Op.getNode(), DAG)) 9613 return Res; 9614 9615 EVT VT = Op.getValueType(); 9616 9617 SDValue LHS = Op.getOperand(0); 9618 BuildVectorSDNode *BVN = 9619 dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode()); 9620 if (!BVN) { 9621 // OR commutes, so try swapping the operands. 9622 LHS = Op.getOperand(1); 9623 BVN = dyn_cast<BuildVectorSDNode>(Op.getOperand(0).getNode()); 9624 } 9625 if (!BVN) 9626 return Op; 9627 9628 APInt DefBits(VT.getSizeInBits(), 0); 9629 APInt UndefBits(VT.getSizeInBits(), 0); 9630 if (resolveBuildVector(BVN, DefBits, UndefBits)) { 9631 SDValue NewOp; 9632 9633 if ((NewOp = tryAdvSIMDModImm32(AArch64ISD::ORRi, Op, DAG, 9634 DefBits, &LHS)) || 9635 (NewOp = tryAdvSIMDModImm16(AArch64ISD::ORRi, Op, DAG, 9636 DefBits, &LHS))) 9637 return NewOp; 9638 9639 if ((NewOp = tryAdvSIMDModImm32(AArch64ISD::ORRi, Op, DAG, 9640 UndefBits, &LHS)) || 9641 (NewOp = tryAdvSIMDModImm16(AArch64ISD::ORRi, Op, DAG, 9642 UndefBits, &LHS))) 9643 return NewOp; 9644 } 9645 9646 // We can always fall back to a non-immediate OR. 9647 return Op; 9648 } 9649 9650 // Normalize the operands of BUILD_VECTOR. The value of constant operands will 9651 // be truncated to fit element width. 9652 static SDValue NormalizeBuildVector(SDValue Op, 9653 SelectionDAG &DAG) { 9654 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!"); 9655 SDLoc dl(Op); 9656 EVT VT = Op.getValueType(); 9657 EVT EltTy= VT.getVectorElementType(); 9658 9659 if (EltTy.isFloatingPoint() || EltTy.getSizeInBits() > 16) 9660 return Op; 9661 9662 SmallVector<SDValue, 16> Ops; 9663 for (SDValue Lane : Op->ops()) { 9664 // For integer vectors, type legalization would have promoted the 9665 // operands already. Otherwise, if Op is a floating-point splat 9666 // (with operands cast to integers), then the only possibilities 9667 // are constants and UNDEFs. 9668 if (auto *CstLane = dyn_cast<ConstantSDNode>(Lane)) { 9669 APInt LowBits(EltTy.getSizeInBits(), 9670 CstLane->getZExtValue()); 9671 Lane = DAG.getConstant(LowBits.getZExtValue(), dl, MVT::i32); 9672 } else if (Lane.getNode()->isUndef()) { 9673 Lane = DAG.getUNDEF(MVT::i32); 9674 } else { 9675 assert(Lane.getValueType() == MVT::i32 && 9676 "Unexpected BUILD_VECTOR operand type"); 9677 } 9678 Ops.push_back(Lane); 9679 } 9680 return DAG.getBuildVector(VT, dl, Ops); 9681 } 9682 9683 static SDValue ConstantBuildVector(SDValue Op, SelectionDAG &DAG) { 9684 EVT VT = Op.getValueType(); 9685 9686 APInt DefBits(VT.getSizeInBits(), 0); 9687 APInt UndefBits(VT.getSizeInBits(), 0); 9688 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode()); 9689 if (resolveBuildVector(BVN, DefBits, UndefBits)) { 9690 SDValue NewOp; 9691 if ((NewOp = tryAdvSIMDModImm64(AArch64ISD::MOVIedit, Op, DAG, DefBits)) || 9692 (NewOp = tryAdvSIMDModImm32(AArch64ISD::MOVIshift, Op, DAG, DefBits)) || 9693 (NewOp = tryAdvSIMDModImm321s(AArch64ISD::MOVImsl, Op, DAG, DefBits)) || 9694 (NewOp = tryAdvSIMDModImm16(AArch64ISD::MOVIshift, Op, DAG, DefBits)) || 9695 (NewOp = tryAdvSIMDModImm8(AArch64ISD::MOVI, Op, DAG, DefBits)) || 9696 (NewOp = tryAdvSIMDModImmFP(AArch64ISD::FMOV, Op, DAG, DefBits))) 9697 return NewOp; 9698 9699 DefBits = ~DefBits; 9700 if ((NewOp = tryAdvSIMDModImm32(AArch64ISD::MVNIshift, Op, DAG, DefBits)) || 9701 (NewOp = tryAdvSIMDModImm321s(AArch64ISD::MVNImsl, Op, DAG, DefBits)) || 9702 (NewOp = tryAdvSIMDModImm16(AArch64ISD::MVNIshift, Op, DAG, DefBits))) 9703 return NewOp; 9704 9705 DefBits = UndefBits; 9706 if ((NewOp = tryAdvSIMDModImm64(AArch64ISD::MOVIedit, Op, DAG, DefBits)) || 9707 (NewOp = tryAdvSIMDModImm32(AArch64ISD::MOVIshift, Op, DAG, DefBits)) || 9708 (NewOp = tryAdvSIMDModImm321s(AArch64ISD::MOVImsl, Op, DAG, DefBits)) || 9709 (NewOp = tryAdvSIMDModImm16(AArch64ISD::MOVIshift, Op, DAG, DefBits)) || 9710 (NewOp = tryAdvSIMDModImm8(AArch64ISD::MOVI, Op, DAG, DefBits)) || 9711 (NewOp = tryAdvSIMDModImmFP(AArch64ISD::FMOV, Op, DAG, DefBits))) 9712 return NewOp; 9713 9714 DefBits = ~UndefBits; 9715 if ((NewOp = tryAdvSIMDModImm32(AArch64ISD::MVNIshift, Op, DAG, DefBits)) || 9716 (NewOp = tryAdvSIMDModImm321s(AArch64ISD::MVNImsl, Op, DAG, DefBits)) || 9717 (NewOp = tryAdvSIMDModImm16(AArch64ISD::MVNIshift, Op, DAG, DefBits))) 9718 return NewOp; 9719 } 9720 9721 return SDValue(); 9722 } 9723 9724 SDValue AArch64TargetLowering::LowerBUILD_VECTOR(SDValue Op, 9725 SelectionDAG &DAG) const { 9726 EVT VT = Op.getValueType(); 9727 9728 // Try to build a simple constant vector. 9729 Op = NormalizeBuildVector(Op, DAG); 9730 if (VT.isInteger()) { 9731 // Certain vector constants, used to express things like logical NOT and 9732 // arithmetic NEG, are passed through unmodified. This allows special 9733 // patterns for these operations to match, which will lower these constants 9734 // to whatever is proven necessary. 9735 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode()); 9736 if (BVN->isConstant()) 9737 if (ConstantSDNode *Const = BVN->getConstantSplatNode()) { 9738 unsigned BitSize = VT.getVectorElementType().getSizeInBits(); 9739 APInt Val(BitSize, 9740 Const->getAPIntValue().zextOrTrunc(BitSize).getZExtValue()); 9741 if (Val.isNullValue() || Val.isAllOnesValue()) 9742 return Op; 9743 } 9744 } 9745 9746 if (SDValue V = ConstantBuildVector(Op, DAG)) 9747 return V; 9748 9749 // Scan through the operands to find some interesting properties we can 9750 // exploit: 9751 // 1) If only one value is used, we can use a DUP, or 9752 // 2) if only the low element is not undef, we can just insert that, or 9753 // 3) if only one constant value is used (w/ some non-constant lanes), 9754 // we can splat the constant value into the whole vector then fill 9755 // in the non-constant lanes. 9756 // 4) FIXME: If different constant values are used, but we can intelligently 9757 // select the values we'll be overwriting for the non-constant 9758 // lanes such that we can directly materialize the vector 9759 // some other way (MOVI, e.g.), we can be sneaky. 9760 // 5) if all operands are EXTRACT_VECTOR_ELT, check for VUZP. 9761 SDLoc dl(Op); 9762 unsigned NumElts = VT.getVectorNumElements(); 9763 bool isOnlyLowElement = true; 9764 bool usesOnlyOneValue = true; 9765 bool usesOnlyOneConstantValue = true; 9766 bool isConstant = true; 9767 bool AllLanesExtractElt = true; 9768 unsigned NumConstantLanes = 0; 9769 unsigned NumDifferentLanes = 0; 9770 unsigned NumUndefLanes = 0; 9771 SDValue Value; 9772 SDValue ConstantValue; 9773 for (unsigned i = 0; i < NumElts; ++i) { 9774 SDValue V = Op.getOperand(i); 9775 if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) 9776 AllLanesExtractElt = false; 9777 if (V.isUndef()) { 9778 ++NumUndefLanes; 9779 continue; 9780 } 9781 if (i > 0) 9782 isOnlyLowElement = false; 9783 if (!isIntOrFPConstant(V)) 9784 isConstant = false; 9785 9786 if (isIntOrFPConstant(V)) { 9787 ++NumConstantLanes; 9788 if (!ConstantValue.getNode()) 9789 ConstantValue = V; 9790 else if (ConstantValue != V) 9791 usesOnlyOneConstantValue = false; 9792 } 9793 9794 if (!Value.getNode()) 9795 Value = V; 9796 else if (V != Value) { 9797 usesOnlyOneValue = false; 9798 ++NumDifferentLanes; 9799 } 9800 } 9801 9802 if (!Value.getNode()) { 9803 LLVM_DEBUG( 9804 dbgs() << "LowerBUILD_VECTOR: value undefined, creating undef node\n"); 9805 return DAG.getUNDEF(VT); 9806 } 9807 9808 // Convert BUILD_VECTOR where all elements but the lowest are undef into 9809 // SCALAR_TO_VECTOR, except for when we have a single-element constant vector 9810 // as SimplifyDemandedBits will just turn that back into BUILD_VECTOR. 9811 if (isOnlyLowElement && !(NumElts == 1 && isIntOrFPConstant(Value))) { 9812 LLVM_DEBUG(dbgs() << "LowerBUILD_VECTOR: only low element used, creating 1 " 9813 "SCALAR_TO_VECTOR node\n"); 9814 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value); 9815 } 9816 9817 if (AllLanesExtractElt) { 9818 SDNode *Vector = nullptr; 9819 bool Even = false; 9820 bool Odd = false; 9821 // Check whether the extract elements match the Even pattern <0,2,4,...> or 9822 // the Odd pattern <1,3,5,...>. 9823 for (unsigned i = 0; i < NumElts; ++i) { 9824 SDValue V = Op.getOperand(i); 9825 const SDNode *N = V.getNode(); 9826 if (!isa<ConstantSDNode>(N->getOperand(1))) 9827 break; 9828 SDValue N0 = N->getOperand(0); 9829 9830 // All elements are extracted from the same vector. 9831 if (!Vector) { 9832 Vector = N0.getNode(); 9833 // Check that the type of EXTRACT_VECTOR_ELT matches the type of 9834 // BUILD_VECTOR. 9835 if (VT.getVectorElementType() != 9836 N0.getValueType().getVectorElementType()) 9837 break; 9838 } else if (Vector != N0.getNode()) { 9839 Odd = false; 9840 Even = false; 9841 break; 9842 } 9843 9844 // Extracted values are either at Even indices <0,2,4,...> or at Odd 9845 // indices <1,3,5,...>. 9846 uint64_t Val = N->getConstantOperandVal(1); 9847 if (Val == 2 * i) { 9848 Even = true; 9849 continue; 9850 } 9851 if (Val - 1 == 2 * i) { 9852 Odd = true; 9853 continue; 9854 } 9855 9856 // Something does not match: abort. 9857 Odd = false; 9858 Even = false; 9859 break; 9860 } 9861 if (Even || Odd) { 9862 SDValue LHS = 9863 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SDValue(Vector, 0), 9864 DAG.getConstant(0, dl, MVT::i64)); 9865 SDValue RHS = 9866 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SDValue(Vector, 0), 9867 DAG.getConstant(NumElts, dl, MVT::i64)); 9868 9869 if (Even && !Odd) 9870 return DAG.getNode(AArch64ISD::UZP1, dl, DAG.getVTList(VT, VT), LHS, 9871 RHS); 9872 if (Odd && !Even) 9873 return DAG.getNode(AArch64ISD::UZP2, dl, DAG.getVTList(VT, VT), LHS, 9874 RHS); 9875 } 9876 } 9877 9878 // Use DUP for non-constant splats. For f32 constant splats, reduce to 9879 // i32 and try again. 9880 if (usesOnlyOneValue) { 9881 if (!isConstant) { 9882 if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT || 9883 Value.getValueType() != VT) { 9884 LLVM_DEBUG( 9885 dbgs() << "LowerBUILD_VECTOR: use DUP for non-constant splats\n"); 9886 return DAG.getNode(AArch64ISD::DUP, dl, VT, Value); 9887 } 9888 9889 // This is actually a DUPLANExx operation, which keeps everything vectory. 9890 9891 SDValue Lane = Value.getOperand(1); 9892 Value = Value.getOperand(0); 9893 if (Value.getValueSizeInBits() == 64) { 9894 LLVM_DEBUG( 9895 dbgs() << "LowerBUILD_VECTOR: DUPLANE works on 128-bit vectors, " 9896 "widening it\n"); 9897 Value = WidenVector(Value, DAG); 9898 } 9899 9900 unsigned Opcode = getDUPLANEOp(VT.getVectorElementType()); 9901 return DAG.getNode(Opcode, dl, VT, Value, Lane); 9902 } 9903 9904 if (VT.getVectorElementType().isFloatingPoint()) { 9905 SmallVector<SDValue, 8> Ops; 9906 EVT EltTy = VT.getVectorElementType(); 9907 assert ((EltTy == MVT::f16 || EltTy == MVT::bf16 || EltTy == MVT::f32 || 9908 EltTy == MVT::f64) && "Unsupported floating-point vector type"); 9909 LLVM_DEBUG( 9910 dbgs() << "LowerBUILD_VECTOR: float constant splats, creating int " 9911 "BITCASTS, and try again\n"); 9912 MVT NewType = MVT::getIntegerVT(EltTy.getSizeInBits()); 9913 for (unsigned i = 0; i < NumElts; ++i) 9914 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, NewType, Op.getOperand(i))); 9915 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), NewType, NumElts); 9916 SDValue Val = DAG.getBuildVector(VecVT, dl, Ops); 9917 LLVM_DEBUG(dbgs() << "LowerBUILD_VECTOR: trying to lower new vector: "; 9918 Val.dump();); 9919 Val = LowerBUILD_VECTOR(Val, DAG); 9920 if (Val.getNode()) 9921 return DAG.getNode(ISD::BITCAST, dl, VT, Val); 9922 } 9923 } 9924 9925 // If we need to insert a small number of different non-constant elements and 9926 // the vector width is sufficiently large, prefer using DUP with the common 9927 // value and INSERT_VECTOR_ELT for the different lanes. If DUP is preferred, 9928 // skip the constant lane handling below. 9929 bool PreferDUPAndInsert = 9930 !isConstant && NumDifferentLanes >= 1 && 9931 NumDifferentLanes < ((NumElts - NumUndefLanes) / 2) && 9932 NumDifferentLanes >= NumConstantLanes; 9933 9934 // If there was only one constant value used and for more than one lane, 9935 // start by splatting that value, then replace the non-constant lanes. This 9936 // is better than the default, which will perform a separate initialization 9937 // for each lane. 9938 if (!PreferDUPAndInsert && NumConstantLanes > 0 && usesOnlyOneConstantValue) { 9939 // Firstly, try to materialize the splat constant. 9940 SDValue Vec = DAG.getSplatBuildVector(VT, dl, ConstantValue), 9941 Val = ConstantBuildVector(Vec, DAG); 9942 if (!Val) { 9943 // Otherwise, materialize the constant and splat it. 9944 Val = DAG.getNode(AArch64ISD::DUP, dl, VT, ConstantValue); 9945 DAG.ReplaceAllUsesWith(Vec.getNode(), &Val); 9946 } 9947 9948 // Now insert the non-constant lanes. 9949 for (unsigned i = 0; i < NumElts; ++i) { 9950 SDValue V = Op.getOperand(i); 9951 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i64); 9952 if (!isIntOrFPConstant(V)) 9953 // Note that type legalization likely mucked about with the VT of the 9954 // source operand, so we may have to convert it here before inserting. 9955 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, V, LaneIdx); 9956 } 9957 return Val; 9958 } 9959 9960 // This will generate a load from the constant pool. 9961 if (isConstant) { 9962 LLVM_DEBUG( 9963 dbgs() << "LowerBUILD_VECTOR: all elements are constant, use default " 9964 "expansion\n"); 9965 return SDValue(); 9966 } 9967 9968 // Empirical tests suggest this is rarely worth it for vectors of length <= 2. 9969 if (NumElts >= 4) { 9970 if (SDValue shuffle = ReconstructShuffle(Op, DAG)) 9971 return shuffle; 9972 } 9973 9974 if (PreferDUPAndInsert) { 9975 // First, build a constant vector with the common element. 9976 SmallVector<SDValue, 8> Ops(NumElts, Value); 9977 SDValue NewVector = LowerBUILD_VECTOR(DAG.getBuildVector(VT, dl, Ops), DAG); 9978 // Next, insert the elements that do not match the common value. 9979 for (unsigned I = 0; I < NumElts; ++I) 9980 if (Op.getOperand(I) != Value) 9981 NewVector = 9982 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, NewVector, 9983 Op.getOperand(I), DAG.getConstant(I, dl, MVT::i64)); 9984 9985 return NewVector; 9986 } 9987 9988 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we 9989 // know the default expansion would otherwise fall back on something even 9990 // worse. For a vector with one or two non-undef values, that's 9991 // scalar_to_vector for the elements followed by a shuffle (provided the 9992 // shuffle is valid for the target) and materialization element by element 9993 // on the stack followed by a load for everything else. 9994 if (!isConstant && !usesOnlyOneValue) { 9995 LLVM_DEBUG( 9996 dbgs() << "LowerBUILD_VECTOR: alternatives failed, creating sequence " 9997 "of INSERT_VECTOR_ELT\n"); 9998 9999 SDValue Vec = DAG.getUNDEF(VT); 10000 SDValue Op0 = Op.getOperand(0); 10001 unsigned i = 0; 10002 10003 // Use SCALAR_TO_VECTOR for lane zero to 10004 // a) Avoid a RMW dependency on the full vector register, and 10005 // b) Allow the register coalescer to fold away the copy if the 10006 // value is already in an S or D register, and we're forced to emit an 10007 // INSERT_SUBREG that we can't fold anywhere. 10008 // 10009 // We also allow types like i8 and i16 which are illegal scalar but legal 10010 // vector element types. After type-legalization the inserted value is 10011 // extended (i32) and it is safe to cast them to the vector type by ignoring 10012 // the upper bits of the lowest lane (e.g. v8i8, v4i16). 10013 if (!Op0.isUndef()) { 10014 LLVM_DEBUG(dbgs() << "Creating node for op0, it is not undefined:\n"); 10015 Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op0); 10016 ++i; 10017 } 10018 LLVM_DEBUG(if (i < NumElts) dbgs() 10019 << "Creating nodes for the other vector elements:\n";); 10020 for (; i < NumElts; ++i) { 10021 SDValue V = Op.getOperand(i); 10022 if (V.isUndef()) 10023 continue; 10024 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i64); 10025 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx); 10026 } 10027 return Vec; 10028 } 10029 10030 LLVM_DEBUG( 10031 dbgs() << "LowerBUILD_VECTOR: use default expansion, failed to find " 10032 "better alternative\n"); 10033 return SDValue(); 10034 } 10035 10036 SDValue AArch64TargetLowering::LowerCONCAT_VECTORS(SDValue Op, 10037 SelectionDAG &DAG) const { 10038 assert(Op.getValueType().isScalableVector() && 10039 isTypeLegal(Op.getValueType()) && 10040 "Expected legal scalable vector type!"); 10041 10042 if (isTypeLegal(Op.getOperand(0).getValueType()) && Op.getNumOperands() == 2) 10043 return Op; 10044 10045 return SDValue(); 10046 } 10047 10048 SDValue AArch64TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, 10049 SelectionDAG &DAG) const { 10050 assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Unknown opcode!"); 10051 10052 if (useSVEForFixedLengthVectorVT(Op.getValueType())) 10053 return LowerFixedLengthInsertVectorElt(Op, DAG); 10054 10055 // Check for non-constant or out of range lane. 10056 EVT VT = Op.getOperand(0).getValueType(); 10057 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(2)); 10058 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements()) 10059 return SDValue(); 10060 10061 10062 // Insertion/extraction are legal for V128 types. 10063 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 || 10064 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 || 10065 VT == MVT::v8f16 || VT == MVT::v8bf16) 10066 return Op; 10067 10068 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 && 10069 VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16 && 10070 VT != MVT::v4bf16) 10071 return SDValue(); 10072 10073 // For V64 types, we perform insertion by expanding the value 10074 // to a V128 type and perform the insertion on that. 10075 SDLoc DL(Op); 10076 SDValue WideVec = WidenVector(Op.getOperand(0), DAG); 10077 EVT WideTy = WideVec.getValueType(); 10078 10079 SDValue Node = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideTy, WideVec, 10080 Op.getOperand(1), Op.getOperand(2)); 10081 // Re-narrow the resultant vector. 10082 return NarrowVector(Node, DAG); 10083 } 10084 10085 SDValue 10086 AArch64TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, 10087 SelectionDAG &DAG) const { 10088 assert(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unknown opcode!"); 10089 EVT VT = Op.getOperand(0).getValueType(); 10090 10091 if (VT.getScalarType() == MVT::i1) { 10092 // We can't directly extract from an SVE predicate; extend it first. 10093 // (This isn't the only possible lowering, but it's straightforward.) 10094 EVT VectorVT = getPromotedVTForPredicate(VT); 10095 SDLoc DL(Op); 10096 SDValue Extend = 10097 DAG.getNode(ISD::ANY_EXTEND, DL, VectorVT, Op.getOperand(0)); 10098 MVT ExtractTy = VectorVT == MVT::nxv2i64 ? MVT::i64 : MVT::i32; 10099 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtractTy, 10100 Extend, Op.getOperand(1)); 10101 return DAG.getAnyExtOrTrunc(Extract, DL, Op.getValueType()); 10102 } 10103 10104 if (useSVEForFixedLengthVectorVT(VT)) 10105 return LowerFixedLengthExtractVectorElt(Op, DAG); 10106 10107 // Check for non-constant or out of range lane. 10108 ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(1)); 10109 if (!CI || CI->getZExtValue() >= VT.getVectorNumElements()) 10110 return SDValue(); 10111 10112 // Insertion/extraction are legal for V128 types. 10113 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 || 10114 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 || 10115 VT == MVT::v8f16 || VT == MVT::v8bf16) 10116 return Op; 10117 10118 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 && 10119 VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16 && 10120 VT != MVT::v4bf16) 10121 return SDValue(); 10122 10123 // For V64 types, we perform extraction by expanding the value 10124 // to a V128 type and perform the extraction on that. 10125 SDLoc DL(Op); 10126 SDValue WideVec = WidenVector(Op.getOperand(0), DAG); 10127 EVT WideTy = WideVec.getValueType(); 10128 10129 EVT ExtrTy = WideTy.getVectorElementType(); 10130 if (ExtrTy == MVT::i16 || ExtrTy == MVT::i8) 10131 ExtrTy = MVT::i32; 10132 10133 // For extractions, we just return the result directly. 10134 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrTy, WideVec, 10135 Op.getOperand(1)); 10136 } 10137 10138 SDValue AArch64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op, 10139 SelectionDAG &DAG) const { 10140 assert(Op.getValueType().isFixedLengthVector() && 10141 "Only cases that extract a fixed length vector are supported!"); 10142 10143 EVT InVT = Op.getOperand(0).getValueType(); 10144 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue(); 10145 unsigned Size = Op.getValueSizeInBits(); 10146 10147 if (InVT.isScalableVector()) { 10148 // This will be matched by custom code during ISelDAGToDAG. 10149 if (Idx == 0 && isPackedVectorType(InVT, DAG)) 10150 return Op; 10151 10152 return SDValue(); 10153 } 10154 10155 // This will get lowered to an appropriate EXTRACT_SUBREG in ISel. 10156 if (Idx == 0 && InVT.getSizeInBits() <= 128) 10157 return Op; 10158 10159 // If this is extracting the upper 64-bits of a 128-bit vector, we match 10160 // that directly. 10161 if (Size == 64 && Idx * InVT.getScalarSizeInBits() == 64 && 10162 InVT.getSizeInBits() == 128) 10163 return Op; 10164 10165 return SDValue(); 10166 } 10167 10168 SDValue AArch64TargetLowering::LowerINSERT_SUBVECTOR(SDValue Op, 10169 SelectionDAG &DAG) const { 10170 assert(Op.getValueType().isScalableVector() && 10171 "Only expect to lower inserts into scalable vectors!"); 10172 10173 EVT InVT = Op.getOperand(1).getValueType(); 10174 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue(); 10175 10176 if (InVT.isScalableVector()) { 10177 SDLoc DL(Op); 10178 EVT VT = Op.getValueType(); 10179 10180 if (!isTypeLegal(VT) || !VT.isInteger()) 10181 return SDValue(); 10182 10183 SDValue Vec0 = Op.getOperand(0); 10184 SDValue Vec1 = Op.getOperand(1); 10185 10186 // Ensure the subvector is half the size of the main vector. 10187 if (VT.getVectorElementCount() != (InVT.getVectorElementCount() * 2)) 10188 return SDValue(); 10189 10190 // Extend elements of smaller vector... 10191 EVT WideVT = InVT.widenIntegerVectorElementType(*(DAG.getContext())); 10192 SDValue ExtVec = DAG.getNode(ISD::ANY_EXTEND, DL, WideVT, Vec1); 10193 10194 if (Idx == 0) { 10195 SDValue HiVec0 = DAG.getNode(AArch64ISD::UUNPKHI, DL, WideVT, Vec0); 10196 return DAG.getNode(AArch64ISD::UZP1, DL, VT, ExtVec, HiVec0); 10197 } else if (Idx == InVT.getVectorMinNumElements()) { 10198 SDValue LoVec0 = DAG.getNode(AArch64ISD::UUNPKLO, DL, WideVT, Vec0); 10199 return DAG.getNode(AArch64ISD::UZP1, DL, VT, LoVec0, ExtVec); 10200 } 10201 10202 return SDValue(); 10203 } 10204 10205 // This will be matched by custom code during ISelDAGToDAG. 10206 if (Idx == 0 && isPackedVectorType(InVT, DAG) && Op.getOperand(0).isUndef()) 10207 return Op; 10208 10209 return SDValue(); 10210 } 10211 10212 SDValue AArch64TargetLowering::LowerDIV(SDValue Op, SelectionDAG &DAG) const { 10213 EVT VT = Op.getValueType(); 10214 10215 if (useSVEForFixedLengthVectorVT(VT, /*OverrideNEON=*/true)) 10216 return LowerFixedLengthVectorIntDivideToSVE(Op, DAG); 10217 10218 assert(VT.isScalableVector() && "Expected a scalable vector."); 10219 10220 bool Signed = Op.getOpcode() == ISD::SDIV; 10221 unsigned PredOpcode = Signed ? AArch64ISD::SDIV_PRED : AArch64ISD::UDIV_PRED; 10222 10223 if (VT == MVT::nxv4i32 || VT == MVT::nxv2i64) 10224 return LowerToPredicatedOp(Op, DAG, PredOpcode); 10225 10226 // SVE doesn't have i8 and i16 DIV operations; widen them to 32-bit 10227 // operations, and truncate the result. 10228 EVT WidenedVT; 10229 if (VT == MVT::nxv16i8) 10230 WidenedVT = MVT::nxv8i16; 10231 else if (VT == MVT::nxv8i16) 10232 WidenedVT = MVT::nxv4i32; 10233 else 10234 llvm_unreachable("Unexpected Custom DIV operation"); 10235 10236 SDLoc dl(Op); 10237 unsigned UnpkLo = Signed ? AArch64ISD::SUNPKLO : AArch64ISD::UUNPKLO; 10238 unsigned UnpkHi = Signed ? AArch64ISD::SUNPKHI : AArch64ISD::UUNPKHI; 10239 SDValue Op0Lo = DAG.getNode(UnpkLo, dl, WidenedVT, Op.getOperand(0)); 10240 SDValue Op1Lo = DAG.getNode(UnpkLo, dl, WidenedVT, Op.getOperand(1)); 10241 SDValue Op0Hi = DAG.getNode(UnpkHi, dl, WidenedVT, Op.getOperand(0)); 10242 SDValue Op1Hi = DAG.getNode(UnpkHi, dl, WidenedVT, Op.getOperand(1)); 10243 SDValue ResultLo = DAG.getNode(Op.getOpcode(), dl, WidenedVT, Op0Lo, Op1Lo); 10244 SDValue ResultHi = DAG.getNode(Op.getOpcode(), dl, WidenedVT, Op0Hi, Op1Hi); 10245 return DAG.getNode(AArch64ISD::UZP1, dl, VT, ResultLo, ResultHi); 10246 } 10247 10248 bool AArch64TargetLowering::isShuffleMaskLegal(ArrayRef<int> M, EVT VT) const { 10249 // Currently no fixed length shuffles that require SVE are legal. 10250 if (useSVEForFixedLengthVectorVT(VT)) 10251 return false; 10252 10253 if (VT.getVectorNumElements() == 4 && 10254 (VT.is128BitVector() || VT.is64BitVector())) { 10255 unsigned PFIndexes[4]; 10256 for (unsigned i = 0; i != 4; ++i) { 10257 if (M[i] < 0) 10258 PFIndexes[i] = 8; 10259 else 10260 PFIndexes[i] = M[i]; 10261 } 10262 10263 // Compute the index in the perfect shuffle table. 10264 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 + 10265 PFIndexes[2] * 9 + PFIndexes[3]; 10266 unsigned PFEntry = PerfectShuffleTable[PFTableIndex]; 10267 unsigned Cost = (PFEntry >> 30); 10268 10269 if (Cost <= 4) 10270 return true; 10271 } 10272 10273 bool DummyBool; 10274 int DummyInt; 10275 unsigned DummyUnsigned; 10276 10277 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) || isREVMask(M, VT, 64) || 10278 isREVMask(M, VT, 32) || isREVMask(M, VT, 16) || 10279 isEXTMask(M, VT, DummyBool, DummyUnsigned) || 10280 // isTBLMask(M, VT) || // FIXME: Port TBL support from ARM. 10281 isTRNMask(M, VT, DummyUnsigned) || isUZPMask(M, VT, DummyUnsigned) || 10282 isZIPMask(M, VT, DummyUnsigned) || 10283 isTRN_v_undef_Mask(M, VT, DummyUnsigned) || 10284 isUZP_v_undef_Mask(M, VT, DummyUnsigned) || 10285 isZIP_v_undef_Mask(M, VT, DummyUnsigned) || 10286 isINSMask(M, VT.getVectorNumElements(), DummyBool, DummyInt) || 10287 isConcatMask(M, VT, VT.getSizeInBits() == 128)); 10288 } 10289 10290 /// getVShiftImm - Check if this is a valid build_vector for the immediate 10291 /// operand of a vector shift operation, where all the elements of the 10292 /// build_vector must have the same constant integer value. 10293 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) { 10294 // Ignore bit_converts. 10295 while (Op.getOpcode() == ISD::BITCAST) 10296 Op = Op.getOperand(0); 10297 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode()); 10298 APInt SplatBits, SplatUndef; 10299 unsigned SplatBitSize; 10300 bool HasAnyUndefs; 10301 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, 10302 HasAnyUndefs, ElementBits) || 10303 SplatBitSize > ElementBits) 10304 return false; 10305 Cnt = SplatBits.getSExtValue(); 10306 return true; 10307 } 10308 10309 /// isVShiftLImm - Check if this is a valid build_vector for the immediate 10310 /// operand of a vector shift left operation. That value must be in the range: 10311 /// 0 <= Value < ElementBits for a left shift; or 10312 /// 0 <= Value <= ElementBits for a long left shift. 10313 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) { 10314 assert(VT.isVector() && "vector shift count is not a vector type"); 10315 int64_t ElementBits = VT.getScalarSizeInBits(); 10316 if (!getVShiftImm(Op, ElementBits, Cnt)) 10317 return false; 10318 return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits); 10319 } 10320 10321 /// isVShiftRImm - Check if this is a valid build_vector for the immediate 10322 /// operand of a vector shift right operation. The value must be in the range: 10323 /// 1 <= Value <= ElementBits for a right shift; or 10324 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, int64_t &Cnt) { 10325 assert(VT.isVector() && "vector shift count is not a vector type"); 10326 int64_t ElementBits = VT.getScalarSizeInBits(); 10327 if (!getVShiftImm(Op, ElementBits, Cnt)) 10328 return false; 10329 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits)); 10330 } 10331 10332 SDValue AArch64TargetLowering::LowerTRUNCATE(SDValue Op, 10333 SelectionDAG &DAG) const { 10334 EVT VT = Op.getValueType(); 10335 10336 if (VT.getScalarType() == MVT::i1) { 10337 // Lower i1 truncate to `(x & 1) != 0`. 10338 SDLoc dl(Op); 10339 EVT OpVT = Op.getOperand(0).getValueType(); 10340 SDValue Zero = DAG.getConstant(0, dl, OpVT); 10341 SDValue One = DAG.getConstant(1, dl, OpVT); 10342 SDValue And = DAG.getNode(ISD::AND, dl, OpVT, Op.getOperand(0), One); 10343 return DAG.getSetCC(dl, VT, And, Zero, ISD::SETNE); 10344 } 10345 10346 if (!VT.isVector() || VT.isScalableVector()) 10347 return SDValue(); 10348 10349 if (useSVEForFixedLengthVectorVT(Op.getOperand(0).getValueType())) 10350 return LowerFixedLengthVectorTruncateToSVE(Op, DAG); 10351 10352 return SDValue(); 10353 } 10354 10355 SDValue AArch64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op, 10356 SelectionDAG &DAG) const { 10357 EVT VT = Op.getValueType(); 10358 SDLoc DL(Op); 10359 int64_t Cnt; 10360 10361 if (!Op.getOperand(1).getValueType().isVector()) 10362 return Op; 10363 unsigned EltSize = VT.getScalarSizeInBits(); 10364 10365 switch (Op.getOpcode()) { 10366 default: 10367 llvm_unreachable("unexpected shift opcode"); 10368 10369 case ISD::SHL: 10370 if (VT.isScalableVector() || useSVEForFixedLengthVectorVT(VT)) 10371 return LowerToPredicatedOp(Op, DAG, AArch64ISD::SHL_PRED); 10372 10373 if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize) 10374 return DAG.getNode(AArch64ISD::VSHL, DL, VT, Op.getOperand(0), 10375 DAG.getConstant(Cnt, DL, MVT::i32)); 10376 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT, 10377 DAG.getConstant(Intrinsic::aarch64_neon_ushl, DL, 10378 MVT::i32), 10379 Op.getOperand(0), Op.getOperand(1)); 10380 case ISD::SRA: 10381 case ISD::SRL: 10382 if (VT.isScalableVector() || useSVEForFixedLengthVectorVT(VT)) { 10383 unsigned Opc = Op.getOpcode() == ISD::SRA ? AArch64ISD::SRA_PRED 10384 : AArch64ISD::SRL_PRED; 10385 return LowerToPredicatedOp(Op, DAG, Opc); 10386 } 10387 10388 // Right shift immediate 10389 if (isVShiftRImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize) { 10390 unsigned Opc = 10391 (Op.getOpcode() == ISD::SRA) ? AArch64ISD::VASHR : AArch64ISD::VLSHR; 10392 return DAG.getNode(Opc, DL, VT, Op.getOperand(0), 10393 DAG.getConstant(Cnt, DL, MVT::i32)); 10394 } 10395 10396 // Right shift register. Note, there is not a shift right register 10397 // instruction, but the shift left register instruction takes a signed 10398 // value, where negative numbers specify a right shift. 10399 unsigned Opc = (Op.getOpcode() == ISD::SRA) ? Intrinsic::aarch64_neon_sshl 10400 : Intrinsic::aarch64_neon_ushl; 10401 // negate the shift amount 10402 SDValue NegShift = DAG.getNode(AArch64ISD::NEG, DL, VT, Op.getOperand(1)); 10403 SDValue NegShiftLeft = 10404 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT, 10405 DAG.getConstant(Opc, DL, MVT::i32), Op.getOperand(0), 10406 NegShift); 10407 return NegShiftLeft; 10408 } 10409 10410 return SDValue(); 10411 } 10412 10413 static SDValue EmitVectorComparison(SDValue LHS, SDValue RHS, 10414 AArch64CC::CondCode CC, bool NoNans, EVT VT, 10415 const SDLoc &dl, SelectionDAG &DAG) { 10416 EVT SrcVT = LHS.getValueType(); 10417 assert(VT.getSizeInBits() == SrcVT.getSizeInBits() && 10418 "function only supposed to emit natural comparisons"); 10419 10420 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode()); 10421 APInt CnstBits(VT.getSizeInBits(), 0); 10422 APInt UndefBits(VT.getSizeInBits(), 0); 10423 bool IsCnst = BVN && resolveBuildVector(BVN, CnstBits, UndefBits); 10424 bool IsZero = IsCnst && (CnstBits == 0); 10425 10426 if (SrcVT.getVectorElementType().isFloatingPoint()) { 10427 switch (CC) { 10428 default: 10429 return SDValue(); 10430 case AArch64CC::NE: { 10431 SDValue Fcmeq; 10432 if (IsZero) 10433 Fcmeq = DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS); 10434 else 10435 Fcmeq = DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS); 10436 return DAG.getNOT(dl, Fcmeq, VT); 10437 } 10438 case AArch64CC::EQ: 10439 if (IsZero) 10440 return DAG.getNode(AArch64ISD::FCMEQz, dl, VT, LHS); 10441 return DAG.getNode(AArch64ISD::FCMEQ, dl, VT, LHS, RHS); 10442 case AArch64CC::GE: 10443 if (IsZero) 10444 return DAG.getNode(AArch64ISD::FCMGEz, dl, VT, LHS); 10445 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, LHS, RHS); 10446 case AArch64CC::GT: 10447 if (IsZero) 10448 return DAG.getNode(AArch64ISD::FCMGTz, dl, VT, LHS); 10449 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, LHS, RHS); 10450 case AArch64CC::LS: 10451 if (IsZero) 10452 return DAG.getNode(AArch64ISD::FCMLEz, dl, VT, LHS); 10453 return DAG.getNode(AArch64ISD::FCMGE, dl, VT, RHS, LHS); 10454 case AArch64CC::LT: 10455 if (!NoNans) 10456 return SDValue(); 10457 // If we ignore NaNs then we can use to the MI implementation. 10458 LLVM_FALLTHROUGH; 10459 case AArch64CC::MI: 10460 if (IsZero) 10461 return DAG.getNode(AArch64ISD::FCMLTz, dl, VT, LHS); 10462 return DAG.getNode(AArch64ISD::FCMGT, dl, VT, RHS, LHS); 10463 } 10464 } 10465 10466 switch (CC) { 10467 default: 10468 return SDValue(); 10469 case AArch64CC::NE: { 10470 SDValue Cmeq; 10471 if (IsZero) 10472 Cmeq = DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS); 10473 else 10474 Cmeq = DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS); 10475 return DAG.getNOT(dl, Cmeq, VT); 10476 } 10477 case AArch64CC::EQ: 10478 if (IsZero) 10479 return DAG.getNode(AArch64ISD::CMEQz, dl, VT, LHS); 10480 return DAG.getNode(AArch64ISD::CMEQ, dl, VT, LHS, RHS); 10481 case AArch64CC::GE: 10482 if (IsZero) 10483 return DAG.getNode(AArch64ISD::CMGEz, dl, VT, LHS); 10484 return DAG.getNode(AArch64ISD::CMGE, dl, VT, LHS, RHS); 10485 case AArch64CC::GT: 10486 if (IsZero) 10487 return DAG.getNode(AArch64ISD::CMGTz, dl, VT, LHS); 10488 return DAG.getNode(AArch64ISD::CMGT, dl, VT, LHS, RHS); 10489 case AArch64CC::LE: 10490 if (IsZero) 10491 return DAG.getNode(AArch64ISD::CMLEz, dl, VT, LHS); 10492 return DAG.getNode(AArch64ISD::CMGE, dl, VT, RHS, LHS); 10493 case AArch64CC::LS: 10494 return DAG.getNode(AArch64ISD::CMHS, dl, VT, RHS, LHS); 10495 case AArch64CC::LO: 10496 return DAG.getNode(AArch64ISD::CMHI, dl, VT, RHS, LHS); 10497 case AArch64CC::LT: 10498 if (IsZero) 10499 return DAG.getNode(AArch64ISD::CMLTz, dl, VT, LHS); 10500 return DAG.getNode(AArch64ISD::CMGT, dl, VT, RHS, LHS); 10501 case AArch64CC::HI: 10502 return DAG.getNode(AArch64ISD::CMHI, dl, VT, LHS, RHS); 10503 case AArch64CC::HS: 10504 return DAG.getNode(AArch64ISD::CMHS, dl, VT, LHS, RHS); 10505 } 10506 } 10507 10508 SDValue AArch64TargetLowering::LowerVSETCC(SDValue Op, 10509 SelectionDAG &DAG) const { 10510 if (Op.getValueType().isScalableVector()) 10511 return LowerToPredicatedOp(Op, DAG, AArch64ISD::SETCC_MERGE_ZERO); 10512 10513 if (useSVEForFixedLengthVectorVT(Op.getOperand(0).getValueType())) 10514 return LowerFixedLengthVectorSetccToSVE(Op, DAG); 10515 10516 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get(); 10517 SDValue LHS = Op.getOperand(0); 10518 SDValue RHS = Op.getOperand(1); 10519 EVT CmpVT = LHS.getValueType().changeVectorElementTypeToInteger(); 10520 SDLoc dl(Op); 10521 10522 if (LHS.getValueType().getVectorElementType().isInteger()) { 10523 assert(LHS.getValueType() == RHS.getValueType()); 10524 AArch64CC::CondCode AArch64CC = changeIntCCToAArch64CC(CC); 10525 SDValue Cmp = 10526 EmitVectorComparison(LHS, RHS, AArch64CC, false, CmpVT, dl, DAG); 10527 return DAG.getSExtOrTrunc(Cmp, dl, Op.getValueType()); 10528 } 10529 10530 const bool FullFP16 = 10531 static_cast<const AArch64Subtarget &>(DAG.getSubtarget()).hasFullFP16(); 10532 10533 // Make v4f16 (only) fcmp operations utilise vector instructions 10534 // v8f16 support will be a litle more complicated 10535 if (!FullFP16 && LHS.getValueType().getVectorElementType() == MVT::f16) { 10536 if (LHS.getValueType().getVectorNumElements() == 4) { 10537 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::v4f32, LHS); 10538 RHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::v4f32, RHS); 10539 SDValue NewSetcc = DAG.getSetCC(dl, MVT::v4i16, LHS, RHS, CC); 10540 DAG.ReplaceAllUsesWith(Op, NewSetcc); 10541 CmpVT = MVT::v4i32; 10542 } else 10543 return SDValue(); 10544 } 10545 10546 assert((!FullFP16 && LHS.getValueType().getVectorElementType() != MVT::f16) || 10547 LHS.getValueType().getVectorElementType() != MVT::f128); 10548 10549 // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally 10550 // clean. Some of them require two branches to implement. 10551 AArch64CC::CondCode CC1, CC2; 10552 bool ShouldInvert; 10553 changeVectorFPCCToAArch64CC(CC, CC1, CC2, ShouldInvert); 10554 10555 bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath; 10556 SDValue Cmp = 10557 EmitVectorComparison(LHS, RHS, CC1, NoNaNs, CmpVT, dl, DAG); 10558 if (!Cmp.getNode()) 10559 return SDValue(); 10560 10561 if (CC2 != AArch64CC::AL) { 10562 SDValue Cmp2 = 10563 EmitVectorComparison(LHS, RHS, CC2, NoNaNs, CmpVT, dl, DAG); 10564 if (!Cmp2.getNode()) 10565 return SDValue(); 10566 10567 Cmp = DAG.getNode(ISD::OR, dl, CmpVT, Cmp, Cmp2); 10568 } 10569 10570 Cmp = DAG.getSExtOrTrunc(Cmp, dl, Op.getValueType()); 10571 10572 if (ShouldInvert) 10573 Cmp = DAG.getNOT(dl, Cmp, Cmp.getValueType()); 10574 10575 return Cmp; 10576 } 10577 10578 static SDValue getReductionSDNode(unsigned Op, SDLoc DL, SDValue ScalarOp, 10579 SelectionDAG &DAG) { 10580 SDValue VecOp = ScalarOp.getOperand(0); 10581 auto Rdx = DAG.getNode(Op, DL, VecOp.getSimpleValueType(), VecOp); 10582 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ScalarOp.getValueType(), Rdx, 10583 DAG.getConstant(0, DL, MVT::i64)); 10584 } 10585 10586 SDValue AArch64TargetLowering::LowerVECREDUCE(SDValue Op, 10587 SelectionDAG &DAG) const { 10588 SDValue Src = Op.getOperand(0); 10589 10590 // Try to lower fixed length reductions to SVE. 10591 EVT SrcVT = Src.getValueType(); 10592 bool OverrideNEON = Op.getOpcode() == ISD::VECREDUCE_AND || 10593 Op.getOpcode() == ISD::VECREDUCE_OR || 10594 Op.getOpcode() == ISD::VECREDUCE_XOR || 10595 Op.getOpcode() == ISD::VECREDUCE_FADD || 10596 (Op.getOpcode() != ISD::VECREDUCE_ADD && 10597 SrcVT.getVectorElementType() == MVT::i64); 10598 if (SrcVT.isScalableVector() || 10599 useSVEForFixedLengthVectorVT(SrcVT, OverrideNEON)) { 10600 10601 if (SrcVT.getVectorElementType() == MVT::i1) 10602 return LowerPredReductionToSVE(Op, DAG); 10603 10604 switch (Op.getOpcode()) { 10605 case ISD::VECREDUCE_ADD: 10606 return LowerReductionToSVE(AArch64ISD::UADDV_PRED, Op, DAG); 10607 case ISD::VECREDUCE_AND: 10608 return LowerReductionToSVE(AArch64ISD::ANDV_PRED, Op, DAG); 10609 case ISD::VECREDUCE_OR: 10610 return LowerReductionToSVE(AArch64ISD::ORV_PRED, Op, DAG); 10611 case ISD::VECREDUCE_SMAX: 10612 return LowerReductionToSVE(AArch64ISD::SMAXV_PRED, Op, DAG); 10613 case ISD::VECREDUCE_SMIN: 10614 return LowerReductionToSVE(AArch64ISD::SMINV_PRED, Op, DAG); 10615 case ISD::VECREDUCE_UMAX: 10616 return LowerReductionToSVE(AArch64ISD::UMAXV_PRED, Op, DAG); 10617 case ISD::VECREDUCE_UMIN: 10618 return LowerReductionToSVE(AArch64ISD::UMINV_PRED, Op, DAG); 10619 case ISD::VECREDUCE_XOR: 10620 return LowerReductionToSVE(AArch64ISD::EORV_PRED, Op, DAG); 10621 case ISD::VECREDUCE_FADD: 10622 return LowerReductionToSVE(AArch64ISD::FADDV_PRED, Op, DAG); 10623 case ISD::VECREDUCE_FMAX: 10624 return LowerReductionToSVE(AArch64ISD::FMAXNMV_PRED, Op, DAG); 10625 case ISD::VECREDUCE_FMIN: 10626 return LowerReductionToSVE(AArch64ISD::FMINNMV_PRED, Op, DAG); 10627 default: 10628 llvm_unreachable("Unhandled fixed length reduction"); 10629 } 10630 } 10631 10632 // Lower NEON reductions. 10633 SDLoc dl(Op); 10634 switch (Op.getOpcode()) { 10635 case ISD::VECREDUCE_ADD: 10636 return getReductionSDNode(AArch64ISD::UADDV, dl, Op, DAG); 10637 case ISD::VECREDUCE_SMAX: 10638 return getReductionSDNode(AArch64ISD::SMAXV, dl, Op, DAG); 10639 case ISD::VECREDUCE_SMIN: 10640 return getReductionSDNode(AArch64ISD::SMINV, dl, Op, DAG); 10641 case ISD::VECREDUCE_UMAX: 10642 return getReductionSDNode(AArch64ISD::UMAXV, dl, Op, DAG); 10643 case ISD::VECREDUCE_UMIN: 10644 return getReductionSDNode(AArch64ISD::UMINV, dl, Op, DAG); 10645 case ISD::VECREDUCE_FMAX: { 10646 return DAG.getNode( 10647 ISD::INTRINSIC_WO_CHAIN, dl, Op.getValueType(), 10648 DAG.getConstant(Intrinsic::aarch64_neon_fmaxnmv, dl, MVT::i32), 10649 Src); 10650 } 10651 case ISD::VECREDUCE_FMIN: { 10652 return DAG.getNode( 10653 ISD::INTRINSIC_WO_CHAIN, dl, Op.getValueType(), 10654 DAG.getConstant(Intrinsic::aarch64_neon_fminnmv, dl, MVT::i32), 10655 Src); 10656 } 10657 default: 10658 llvm_unreachable("Unhandled reduction"); 10659 } 10660 } 10661 10662 SDValue AArch64TargetLowering::LowerATOMIC_LOAD_SUB(SDValue Op, 10663 SelectionDAG &DAG) const { 10664 auto &Subtarget = static_cast<const AArch64Subtarget &>(DAG.getSubtarget()); 10665 if (!Subtarget.hasLSE() && !Subtarget.outlineAtomics()) 10666 return SDValue(); 10667 10668 // LSE has an atomic load-add instruction, but not a load-sub. 10669 SDLoc dl(Op); 10670 MVT VT = Op.getSimpleValueType(); 10671 SDValue RHS = Op.getOperand(2); 10672 AtomicSDNode *AN = cast<AtomicSDNode>(Op.getNode()); 10673 RHS = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, dl, VT), RHS); 10674 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl, AN->getMemoryVT(), 10675 Op.getOperand(0), Op.getOperand(1), RHS, 10676 AN->getMemOperand()); 10677 } 10678 10679 SDValue AArch64TargetLowering::LowerATOMIC_LOAD_AND(SDValue Op, 10680 SelectionDAG &DAG) const { 10681 auto &Subtarget = static_cast<const AArch64Subtarget &>(DAG.getSubtarget()); 10682 if (!Subtarget.hasLSE() && !Subtarget.outlineAtomics()) 10683 return SDValue(); 10684 10685 // LSE has an atomic load-clear instruction, but not a load-and. 10686 SDLoc dl(Op); 10687 MVT VT = Op.getSimpleValueType(); 10688 SDValue RHS = Op.getOperand(2); 10689 AtomicSDNode *AN = cast<AtomicSDNode>(Op.getNode()); 10690 RHS = DAG.getNode(ISD::XOR, dl, VT, DAG.getConstant(-1ULL, dl, VT), RHS); 10691 return DAG.getAtomic(ISD::ATOMIC_LOAD_CLR, dl, AN->getMemoryVT(), 10692 Op.getOperand(0), Op.getOperand(1), RHS, 10693 AN->getMemOperand()); 10694 } 10695 10696 SDValue AArch64TargetLowering::LowerWindowsDYNAMIC_STACKALLOC( 10697 SDValue Op, SDValue Chain, SDValue &Size, SelectionDAG &DAG) const { 10698 SDLoc dl(Op); 10699 EVT PtrVT = getPointerTy(DAG.getDataLayout()); 10700 SDValue Callee = DAG.getTargetExternalSymbol("__chkstk", PtrVT, 0); 10701 10702 const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo(); 10703 const uint32_t *Mask = TRI->getWindowsStackProbePreservedMask(); 10704 if (Subtarget->hasCustomCallingConv()) 10705 TRI->UpdateCustomCallPreservedMask(DAG.getMachineFunction(), &Mask); 10706 10707 Size = DAG.getNode(ISD::SRL, dl, MVT::i64, Size, 10708 DAG.getConstant(4, dl, MVT::i64)); 10709 Chain = DAG.getCopyToReg(Chain, dl, AArch64::X15, Size, SDValue()); 10710 Chain = 10711 DAG.getNode(AArch64ISD::CALL, dl, DAG.getVTList(MVT::Other, MVT::Glue), 10712 Chain, Callee, DAG.getRegister(AArch64::X15, MVT::i64), 10713 DAG.getRegisterMask(Mask), Chain.getValue(1)); 10714 // To match the actual intent better, we should read the output from X15 here 10715 // again (instead of potentially spilling it to the stack), but rereading Size 10716 // from X15 here doesn't work at -O0, since it thinks that X15 is undefined 10717 // here. 10718 10719 Size = DAG.getNode(ISD::SHL, dl, MVT::i64, Size, 10720 DAG.getConstant(4, dl, MVT::i64)); 10721 return Chain; 10722 } 10723 10724 SDValue 10725 AArch64TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, 10726 SelectionDAG &DAG) const { 10727 assert(Subtarget->isTargetWindows() && 10728 "Only Windows alloca probing supported"); 10729 SDLoc dl(Op); 10730 // Get the inputs. 10731 SDNode *Node = Op.getNode(); 10732 SDValue Chain = Op.getOperand(0); 10733 SDValue Size = Op.getOperand(1); 10734 MaybeAlign Align = 10735 cast<ConstantSDNode>(Op.getOperand(2))->getMaybeAlignValue(); 10736 EVT VT = Node->getValueType(0); 10737 10738 if (DAG.getMachineFunction().getFunction().hasFnAttribute( 10739 "no-stack-arg-probe")) { 10740 SDValue SP = DAG.getCopyFromReg(Chain, dl, AArch64::SP, MVT::i64); 10741 Chain = SP.getValue(1); 10742 SP = DAG.getNode(ISD::SUB, dl, MVT::i64, SP, Size); 10743 if (Align) 10744 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0), 10745 DAG.getConstant(-(uint64_t)Align->value(), dl, VT)); 10746 Chain = DAG.getCopyToReg(Chain, dl, AArch64::SP, SP); 10747 SDValue Ops[2] = {SP, Chain}; 10748 return DAG.getMergeValues(Ops, dl); 10749 } 10750 10751 Chain = DAG.getCALLSEQ_START(Chain, 0, 0, dl); 10752 10753 Chain = LowerWindowsDYNAMIC_STACKALLOC(Op, Chain, Size, DAG); 10754 10755 SDValue SP = DAG.getCopyFromReg(Chain, dl, AArch64::SP, MVT::i64); 10756 Chain = SP.getValue(1); 10757 SP = DAG.getNode(ISD::SUB, dl, MVT::i64, SP, Size); 10758 if (Align) 10759 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0), 10760 DAG.getConstant(-(uint64_t)Align->value(), dl, VT)); 10761 Chain = DAG.getCopyToReg(Chain, dl, AArch64::SP, SP); 10762 10763 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true), 10764 DAG.getIntPtrConstant(0, dl, true), SDValue(), dl); 10765 10766 SDValue Ops[2] = {SP, Chain}; 10767 return DAG.getMergeValues(Ops, dl); 10768 } 10769 10770 SDValue AArch64TargetLowering::LowerVSCALE(SDValue Op, 10771 SelectionDAG &DAG) const { 10772 EVT VT = Op.getValueType(); 10773 assert(VT != MVT::i64 && "Expected illegal VSCALE node"); 10774 10775 SDLoc DL(Op); 10776 APInt MulImm = cast<ConstantSDNode>(Op.getOperand(0))->getAPIntValue(); 10777 return DAG.getZExtOrTrunc(DAG.getVScale(DL, MVT::i64, MulImm.sextOrSelf(64)), 10778 DL, VT); 10779 } 10780 10781 /// Set the IntrinsicInfo for the `aarch64_sve_st<N>` intrinsics. 10782 template <unsigned NumVecs> 10783 static bool 10784 setInfoSVEStN(const AArch64TargetLowering &TLI, const DataLayout &DL, 10785 AArch64TargetLowering::IntrinsicInfo &Info, const CallInst &CI) { 10786 Info.opc = ISD::INTRINSIC_VOID; 10787 // Retrieve EC from first vector argument. 10788 const EVT VT = TLI.getMemValueType(DL, CI.getArgOperand(0)->getType()); 10789 ElementCount EC = VT.getVectorElementCount(); 10790 #ifndef NDEBUG 10791 // Check the assumption that all input vectors are the same type. 10792 for (unsigned I = 0; I < NumVecs; ++I) 10793 assert(VT == TLI.getMemValueType(DL, CI.getArgOperand(I)->getType()) && 10794 "Invalid type."); 10795 #endif 10796 // memVT is `NumVecs * VT`. 10797 Info.memVT = EVT::getVectorVT(CI.getType()->getContext(), VT.getScalarType(), 10798 EC * NumVecs); 10799 Info.ptrVal = CI.getArgOperand(CI.getNumArgOperands() - 1); 10800 Info.offset = 0; 10801 Info.align.reset(); 10802 Info.flags = MachineMemOperand::MOStore; 10803 return true; 10804 } 10805 10806 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as 10807 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment 10808 /// specified in the intrinsic calls. 10809 bool AArch64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info, 10810 const CallInst &I, 10811 MachineFunction &MF, 10812 unsigned Intrinsic) const { 10813 auto &DL = I.getModule()->getDataLayout(); 10814 switch (Intrinsic) { 10815 case Intrinsic::aarch64_sve_st2: 10816 return setInfoSVEStN<2>(*this, DL, Info, I); 10817 case Intrinsic::aarch64_sve_st3: 10818 return setInfoSVEStN<3>(*this, DL, Info, I); 10819 case Intrinsic::aarch64_sve_st4: 10820 return setInfoSVEStN<4>(*this, DL, Info, I); 10821 case Intrinsic::aarch64_neon_ld2: 10822 case Intrinsic::aarch64_neon_ld3: 10823 case Intrinsic::aarch64_neon_ld4: 10824 case Intrinsic::aarch64_neon_ld1x2: 10825 case Intrinsic::aarch64_neon_ld1x3: 10826 case Intrinsic::aarch64_neon_ld1x4: 10827 case Intrinsic::aarch64_neon_ld2lane: 10828 case Intrinsic::aarch64_neon_ld3lane: 10829 case Intrinsic::aarch64_neon_ld4lane: 10830 case Intrinsic::aarch64_neon_ld2r: 10831 case Intrinsic::aarch64_neon_ld3r: 10832 case Intrinsic::aarch64_neon_ld4r: { 10833 Info.opc = ISD::INTRINSIC_W_CHAIN; 10834 // Conservatively set memVT to the entire set of vectors loaded. 10835 uint64_t NumElts = DL.getTypeSizeInBits(I.getType()) / 64; 10836 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts); 10837 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1); 10838 Info.offset = 0; 10839 Info.align.reset(); 10840 // volatile loads with NEON intrinsics not supported 10841 Info.flags = MachineMemOperand::MOLoad; 10842 return true; 10843 } 10844 case Intrinsic::aarch64_neon_st2: 10845 case Intrinsic::aarch64_neon_st3: 10846 case Intrinsic::aarch64_neon_st4: 10847 case Intrinsic::aarch64_neon_st1x2: 10848 case Intrinsic::aarch64_neon_st1x3: 10849 case Intrinsic::aarch64_neon_st1x4: 10850 case Intrinsic::aarch64_neon_st2lane: 10851 case Intrinsic::aarch64_neon_st3lane: 10852 case Intrinsic::aarch64_neon_st4lane: { 10853 Info.opc = ISD::INTRINSIC_VOID; 10854 // Conservatively set memVT to the entire set of vectors stored. 10855 unsigned NumElts = 0; 10856 for (unsigned ArgI = 0, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) { 10857 Type *ArgTy = I.getArgOperand(ArgI)->getType(); 10858 if (!ArgTy->isVectorTy()) 10859 break; 10860 NumElts += DL.getTypeSizeInBits(ArgTy) / 64; 10861 } 10862 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts); 10863 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1); 10864 Info.offset = 0; 10865 Info.align.reset(); 10866 // volatile stores with NEON intrinsics not supported 10867 Info.flags = MachineMemOperand::MOStore; 10868 return true; 10869 } 10870 case Intrinsic::aarch64_ldaxr: 10871 case Intrinsic::aarch64_ldxr: { 10872 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType()); 10873 Info.opc = ISD::INTRINSIC_W_CHAIN; 10874 Info.memVT = MVT::getVT(PtrTy->getElementType()); 10875 Info.ptrVal = I.getArgOperand(0); 10876 Info.offset = 0; 10877 Info.align = DL.getABITypeAlign(PtrTy->getElementType()); 10878 Info.flags = MachineMemOperand::MOLoad | MachineMemOperand::MOVolatile; 10879 return true; 10880 } 10881 case Intrinsic::aarch64_stlxr: 10882 case Intrinsic::aarch64_stxr: { 10883 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType()); 10884 Info.opc = ISD::INTRINSIC_W_CHAIN; 10885 Info.memVT = MVT::getVT(PtrTy->getElementType()); 10886 Info.ptrVal = I.getArgOperand(1); 10887 Info.offset = 0; 10888 Info.align = DL.getABITypeAlign(PtrTy->getElementType()); 10889 Info.flags = MachineMemOperand::MOStore | MachineMemOperand::MOVolatile; 10890 return true; 10891 } 10892 case Intrinsic::aarch64_ldaxp: 10893 case Intrinsic::aarch64_ldxp: 10894 Info.opc = ISD::INTRINSIC_W_CHAIN; 10895 Info.memVT = MVT::i128; 10896 Info.ptrVal = I.getArgOperand(0); 10897 Info.offset = 0; 10898 Info.align = Align(16); 10899 Info.flags = MachineMemOperand::MOLoad | MachineMemOperand::MOVolatile; 10900 return true; 10901 case Intrinsic::aarch64_stlxp: 10902 case Intrinsic::aarch64_stxp: 10903 Info.opc = ISD::INTRINSIC_W_CHAIN; 10904 Info.memVT = MVT::i128; 10905 Info.ptrVal = I.getArgOperand(2); 10906 Info.offset = 0; 10907 Info.align = Align(16); 10908 Info.flags = MachineMemOperand::MOStore | MachineMemOperand::MOVolatile; 10909 return true; 10910 case Intrinsic::aarch64_sve_ldnt1: { 10911 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType()); 10912 Info.opc = ISD::INTRINSIC_W_CHAIN; 10913 Info.memVT = MVT::getVT(I.getType()); 10914 Info.ptrVal = I.getArgOperand(1); 10915 Info.offset = 0; 10916 Info.align = DL.getABITypeAlign(PtrTy->getElementType()); 10917 Info.flags = MachineMemOperand::MOLoad; 10918 if (Intrinsic == Intrinsic::aarch64_sve_ldnt1) 10919 Info.flags |= MachineMemOperand::MONonTemporal; 10920 return true; 10921 } 10922 case Intrinsic::aarch64_sve_stnt1: { 10923 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(2)->getType()); 10924 Info.opc = ISD::INTRINSIC_W_CHAIN; 10925 Info.memVT = MVT::getVT(I.getOperand(0)->getType()); 10926 Info.ptrVal = I.getArgOperand(2); 10927 Info.offset = 0; 10928 Info.align = DL.getABITypeAlign(PtrTy->getElementType()); 10929 Info.flags = MachineMemOperand::MOStore; 10930 if (Intrinsic == Intrinsic::aarch64_sve_stnt1) 10931 Info.flags |= MachineMemOperand::MONonTemporal; 10932 return true; 10933 } 10934 default: 10935 break; 10936 } 10937 10938 return false; 10939 } 10940 10941 bool AArch64TargetLowering::shouldReduceLoadWidth(SDNode *Load, 10942 ISD::LoadExtType ExtTy, 10943 EVT NewVT) const { 10944 // TODO: This may be worth removing. Check regression tests for diffs. 10945 if (!TargetLoweringBase::shouldReduceLoadWidth(Load, ExtTy, NewVT)) 10946 return false; 10947 10948 // If we're reducing the load width in order to avoid having to use an extra 10949 // instruction to do extension then it's probably a good idea. 10950 if (ExtTy != ISD::NON_EXTLOAD) 10951 return true; 10952 // Don't reduce load width if it would prevent us from combining a shift into 10953 // the offset. 10954 MemSDNode *Mem = dyn_cast<MemSDNode>(Load); 10955 assert(Mem); 10956 const SDValue &Base = Mem->getBasePtr(); 10957 if (Base.getOpcode() == ISD::ADD && 10958 Base.getOperand(1).getOpcode() == ISD::SHL && 10959 Base.getOperand(1).hasOneUse() && 10960 Base.getOperand(1).getOperand(1).getOpcode() == ISD::Constant) { 10961 // The shift can be combined if it matches the size of the value being 10962 // loaded (and so reducing the width would make it not match). 10963 uint64_t ShiftAmount = Base.getOperand(1).getConstantOperandVal(1); 10964 uint64_t LoadBytes = Mem->getMemoryVT().getSizeInBits()/8; 10965 if (ShiftAmount == Log2_32(LoadBytes)) 10966 return false; 10967 } 10968 // We have no reason to disallow reducing the load width, so allow it. 10969 return true; 10970 } 10971 10972 // Truncations from 64-bit GPR to 32-bit GPR is free. 10973 bool AArch64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const { 10974 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy()) 10975 return false; 10976 uint64_t NumBits1 = Ty1->getPrimitiveSizeInBits().getFixedSize(); 10977 uint64_t NumBits2 = Ty2->getPrimitiveSizeInBits().getFixedSize(); 10978 return NumBits1 > NumBits2; 10979 } 10980 bool AArch64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const { 10981 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger()) 10982 return false; 10983 uint64_t NumBits1 = VT1.getFixedSizeInBits(); 10984 uint64_t NumBits2 = VT2.getFixedSizeInBits(); 10985 return NumBits1 > NumBits2; 10986 } 10987 10988 /// Check if it is profitable to hoist instruction in then/else to if. 10989 /// Not profitable if I and it's user can form a FMA instruction 10990 /// because we prefer FMSUB/FMADD. 10991 bool AArch64TargetLowering::isProfitableToHoist(Instruction *I) const { 10992 if (I->getOpcode() != Instruction::FMul) 10993 return true; 10994 10995 if (!I->hasOneUse()) 10996 return true; 10997 10998 Instruction *User = I->user_back(); 10999 11000 if (User && 11001 !(User->getOpcode() == Instruction::FSub || 11002 User->getOpcode() == Instruction::FAdd)) 11003 return true; 11004 11005 const TargetOptions &Options = getTargetMachine().Options; 11006 const Function *F = I->getFunction(); 11007 const DataLayout &DL = F->getParent()->getDataLayout(); 11008 Type *Ty = User->getOperand(0)->getType(); 11009 11010 return !(isFMAFasterThanFMulAndFAdd(*F, Ty) && 11011 isOperationLegalOrCustom(ISD::FMA, getValueType(DL, Ty)) && 11012 (Options.AllowFPOpFusion == FPOpFusion::Fast || 11013 Options.UnsafeFPMath)); 11014 } 11015 11016 // All 32-bit GPR operations implicitly zero the high-half of the corresponding 11017 // 64-bit GPR. 11018 bool AArch64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const { 11019 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy()) 11020 return false; 11021 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits(); 11022 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits(); 11023 return NumBits1 == 32 && NumBits2 == 64; 11024 } 11025 bool AArch64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const { 11026 if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger()) 11027 return false; 11028 unsigned NumBits1 = VT1.getSizeInBits(); 11029 unsigned NumBits2 = VT2.getSizeInBits(); 11030 return NumBits1 == 32 && NumBits2 == 64; 11031 } 11032 11033 bool AArch64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const { 11034 EVT VT1 = Val.getValueType(); 11035 if (isZExtFree(VT1, VT2)) { 11036 return true; 11037 } 11038 11039 if (Val.getOpcode() != ISD::LOAD) 11040 return false; 11041 11042 // 8-, 16-, and 32-bit integer loads all implicitly zero-extend. 11043 return (VT1.isSimple() && !VT1.isVector() && VT1.isInteger() && 11044 VT2.isSimple() && !VT2.isVector() && VT2.isInteger() && 11045 VT1.getSizeInBits() <= 32); 11046 } 11047 11048 bool AArch64TargetLowering::isExtFreeImpl(const Instruction *Ext) const { 11049 if (isa<FPExtInst>(Ext)) 11050 return false; 11051 11052 // Vector types are not free. 11053 if (Ext->getType()->isVectorTy()) 11054 return false; 11055 11056 for (const Use &U : Ext->uses()) { 11057 // The extension is free if we can fold it with a left shift in an 11058 // addressing mode or an arithmetic operation: add, sub, and cmp. 11059 11060 // Is there a shift? 11061 const Instruction *Instr = cast<Instruction>(U.getUser()); 11062 11063 // Is this a constant shift? 11064 switch (Instr->getOpcode()) { 11065 case Instruction::Shl: 11066 if (!isa<ConstantInt>(Instr->getOperand(1))) 11067 return false; 11068 break; 11069 case Instruction::GetElementPtr: { 11070 gep_type_iterator GTI = gep_type_begin(Instr); 11071 auto &DL = Ext->getModule()->getDataLayout(); 11072 std::advance(GTI, U.getOperandNo()-1); 11073 Type *IdxTy = GTI.getIndexedType(); 11074 // This extension will end up with a shift because of the scaling factor. 11075 // 8-bit sized types have a scaling factor of 1, thus a shift amount of 0. 11076 // Get the shift amount based on the scaling factor: 11077 // log2(sizeof(IdxTy)) - log2(8). 11078 uint64_t ShiftAmt = 11079 countTrailingZeros(DL.getTypeStoreSizeInBits(IdxTy).getFixedSize()) - 3; 11080 // Is the constant foldable in the shift of the addressing mode? 11081 // I.e., shift amount is between 1 and 4 inclusive. 11082 if (ShiftAmt == 0 || ShiftAmt > 4) 11083 return false; 11084 break; 11085 } 11086 case Instruction::Trunc: 11087 // Check if this is a noop. 11088 // trunc(sext ty1 to ty2) to ty1. 11089 if (Instr->getType() == Ext->getOperand(0)->getType()) 11090 continue; 11091 LLVM_FALLTHROUGH; 11092 default: 11093 return false; 11094 } 11095 11096 // At this point we can use the bfm family, so this extension is free 11097 // for that use. 11098 } 11099 return true; 11100 } 11101 11102 /// Check if both Op1 and Op2 are shufflevector extracts of either the lower 11103 /// or upper half of the vector elements. 11104 static bool areExtractShuffleVectors(Value *Op1, Value *Op2) { 11105 auto areTypesHalfed = [](Value *FullV, Value *HalfV) { 11106 auto *FullTy = FullV->getType(); 11107 auto *HalfTy = HalfV->getType(); 11108 return FullTy->getPrimitiveSizeInBits().getFixedSize() == 11109 2 * HalfTy->getPrimitiveSizeInBits().getFixedSize(); 11110 }; 11111 11112 auto extractHalf = [](Value *FullV, Value *HalfV) { 11113 auto *FullVT = cast<FixedVectorType>(FullV->getType()); 11114 auto *HalfVT = cast<FixedVectorType>(HalfV->getType()); 11115 return FullVT->getNumElements() == 2 * HalfVT->getNumElements(); 11116 }; 11117 11118 ArrayRef<int> M1, M2; 11119 Value *S1Op1, *S2Op1; 11120 if (!match(Op1, m_Shuffle(m_Value(S1Op1), m_Undef(), m_Mask(M1))) || 11121 !match(Op2, m_Shuffle(m_Value(S2Op1), m_Undef(), m_Mask(M2)))) 11122 return false; 11123 11124 // Check that the operands are half as wide as the result and we extract 11125 // half of the elements of the input vectors. 11126 if (!areTypesHalfed(S1Op1, Op1) || !areTypesHalfed(S2Op1, Op2) || 11127 !extractHalf(S1Op1, Op1) || !extractHalf(S2Op1, Op2)) 11128 return false; 11129 11130 // Check the mask extracts either the lower or upper half of vector 11131 // elements. 11132 int M1Start = -1; 11133 int M2Start = -1; 11134 int NumElements = cast<FixedVectorType>(Op1->getType())->getNumElements() * 2; 11135 if (!ShuffleVectorInst::isExtractSubvectorMask(M1, NumElements, M1Start) || 11136 !ShuffleVectorInst::isExtractSubvectorMask(M2, NumElements, M2Start) || 11137 M1Start != M2Start || (M1Start != 0 && M2Start != (NumElements / 2))) 11138 return false; 11139 11140 return true; 11141 } 11142 11143 /// Check if Ext1 and Ext2 are extends of the same type, doubling the bitwidth 11144 /// of the vector elements. 11145 static bool areExtractExts(Value *Ext1, Value *Ext2) { 11146 auto areExtDoubled = [](Instruction *Ext) { 11147 return Ext->getType()->getScalarSizeInBits() == 11148 2 * Ext->getOperand(0)->getType()->getScalarSizeInBits(); 11149 }; 11150 11151 if (!match(Ext1, m_ZExtOrSExt(m_Value())) || 11152 !match(Ext2, m_ZExtOrSExt(m_Value())) || 11153 !areExtDoubled(cast<Instruction>(Ext1)) || 11154 !areExtDoubled(cast<Instruction>(Ext2))) 11155 return false; 11156 11157 return true; 11158 } 11159 11160 /// Check if Op could be used with vmull_high_p64 intrinsic. 11161 static bool isOperandOfVmullHighP64(Value *Op) { 11162 Value *VectorOperand = nullptr; 11163 ConstantInt *ElementIndex = nullptr; 11164 return match(Op, m_ExtractElt(m_Value(VectorOperand), 11165 m_ConstantInt(ElementIndex))) && 11166 ElementIndex->getValue() == 1 && 11167 isa<FixedVectorType>(VectorOperand->getType()) && 11168 cast<FixedVectorType>(VectorOperand->getType())->getNumElements() == 2; 11169 } 11170 11171 /// Check if Op1 and Op2 could be used with vmull_high_p64 intrinsic. 11172 static bool areOperandsOfVmullHighP64(Value *Op1, Value *Op2) { 11173 return isOperandOfVmullHighP64(Op1) && isOperandOfVmullHighP64(Op2); 11174 } 11175 11176 /// Check if sinking \p I's operands to I's basic block is profitable, because 11177 /// the operands can be folded into a target instruction, e.g. 11178 /// shufflevectors extracts and/or sext/zext can be folded into (u,s)subl(2). 11179 bool AArch64TargetLowering::shouldSinkOperands( 11180 Instruction *I, SmallVectorImpl<Use *> &Ops) const { 11181 if (!I->getType()->isVectorTy()) 11182 return false; 11183 11184 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 11185 switch (II->getIntrinsicID()) { 11186 case Intrinsic::aarch64_neon_umull: 11187 if (!areExtractShuffleVectors(II->getOperand(0), II->getOperand(1))) 11188 return false; 11189 Ops.push_back(&II->getOperandUse(0)); 11190 Ops.push_back(&II->getOperandUse(1)); 11191 return true; 11192 11193 case Intrinsic::aarch64_neon_pmull64: 11194 if (!areOperandsOfVmullHighP64(II->getArgOperand(0), 11195 II->getArgOperand(1))) 11196 return false; 11197 Ops.push_back(&II->getArgOperandUse(0)); 11198 Ops.push_back(&II->getArgOperandUse(1)); 11199 return true; 11200 11201 default: 11202 return false; 11203 } 11204 } 11205 11206 switch (I->getOpcode()) { 11207 case Instruction::Sub: 11208 case Instruction::Add: { 11209 if (!areExtractExts(I->getOperand(0), I->getOperand(1))) 11210 return false; 11211 11212 // If the exts' operands extract either the lower or upper elements, we 11213 // can sink them too. 11214 auto Ext1 = cast<Instruction>(I->getOperand(0)); 11215 auto Ext2 = cast<Instruction>(I->getOperand(1)); 11216 if (areExtractShuffleVectors(Ext1, Ext2)) { 11217 Ops.push_back(&Ext1->getOperandUse(0)); 11218 Ops.push_back(&Ext2->getOperandUse(0)); 11219 } 11220 11221 Ops.push_back(&I->getOperandUse(0)); 11222 Ops.push_back(&I->getOperandUse(1)); 11223 11224 return true; 11225 } 11226 case Instruction::Mul: { 11227 bool IsProfitable = false; 11228 for (auto &Op : I->operands()) { 11229 // Make sure we are not already sinking this operand 11230 if (any_of(Ops, [&](Use *U) { return U->get() == Op; })) 11231 continue; 11232 11233 ShuffleVectorInst *Shuffle = dyn_cast<ShuffleVectorInst>(Op); 11234 if (!Shuffle || !Shuffle->isZeroEltSplat()) 11235 continue; 11236 11237 Value *ShuffleOperand = Shuffle->getOperand(0); 11238 InsertElementInst *Insert = dyn_cast<InsertElementInst>(ShuffleOperand); 11239 if (!Insert) 11240 continue; 11241 11242 Instruction *OperandInstr = dyn_cast<Instruction>(Insert->getOperand(1)); 11243 if (!OperandInstr) 11244 continue; 11245 11246 ConstantInt *ElementConstant = 11247 dyn_cast<ConstantInt>(Insert->getOperand(2)); 11248 // Check that the insertelement is inserting into element 0 11249 if (!ElementConstant || ElementConstant->getZExtValue() != 0) 11250 continue; 11251 11252 unsigned Opcode = OperandInstr->getOpcode(); 11253 if (Opcode != Instruction::SExt && Opcode != Instruction::ZExt) 11254 continue; 11255 11256 Ops.push_back(&Shuffle->getOperandUse(0)); 11257 Ops.push_back(&Op); 11258 IsProfitable = true; 11259 } 11260 11261 return IsProfitable; 11262 } 11263 default: 11264 return false; 11265 } 11266 return false; 11267 } 11268 11269 bool AArch64TargetLowering::hasPairedLoad(EVT LoadedType, 11270 Align &RequiredAligment) const { 11271 if (!LoadedType.isSimple() || 11272 (!LoadedType.isInteger() && !LoadedType.isFloatingPoint())) 11273 return false; 11274 // Cyclone supports unaligned accesses. 11275 RequiredAligment = Align(1); 11276 unsigned NumBits = LoadedType.getSizeInBits(); 11277 return NumBits == 32 || NumBits == 64; 11278 } 11279 11280 /// A helper function for determining the number of interleaved accesses we 11281 /// will generate when lowering accesses of the given type. 11282 unsigned 11283 AArch64TargetLowering::getNumInterleavedAccesses(VectorType *VecTy, 11284 const DataLayout &DL) const { 11285 return (DL.getTypeSizeInBits(VecTy) + 127) / 128; 11286 } 11287 11288 MachineMemOperand::Flags 11289 AArch64TargetLowering::getTargetMMOFlags(const Instruction &I) const { 11290 if (Subtarget->getProcFamily() == AArch64Subtarget::Falkor && 11291 I.getMetadata(FALKOR_STRIDED_ACCESS_MD) != nullptr) 11292 return MOStridedAccess; 11293 return MachineMemOperand::MONone; 11294 } 11295 11296 bool AArch64TargetLowering::isLegalInterleavedAccessType( 11297 VectorType *VecTy, const DataLayout &DL) const { 11298 11299 unsigned VecSize = DL.getTypeSizeInBits(VecTy); 11300 unsigned ElSize = DL.getTypeSizeInBits(VecTy->getElementType()); 11301 11302 // Ensure the number of vector elements is greater than 1. 11303 if (cast<FixedVectorType>(VecTy)->getNumElements() < 2) 11304 return false; 11305 11306 // Ensure the element type is legal. 11307 if (ElSize != 8 && ElSize != 16 && ElSize != 32 && ElSize != 64) 11308 return false; 11309 11310 // Ensure the total vector size is 64 or a multiple of 128. Types larger than 11311 // 128 will be split into multiple interleaved accesses. 11312 return VecSize == 64 || VecSize % 128 == 0; 11313 } 11314 11315 /// Lower an interleaved load into a ldN intrinsic. 11316 /// 11317 /// E.g. Lower an interleaved load (Factor = 2): 11318 /// %wide.vec = load <8 x i32>, <8 x i32>* %ptr 11319 /// %v0 = shuffle %wide.vec, undef, <0, 2, 4, 6> ; Extract even elements 11320 /// %v1 = shuffle %wide.vec, undef, <1, 3, 5, 7> ; Extract odd elements 11321 /// 11322 /// Into: 11323 /// %ld2 = { <4 x i32>, <4 x i32> } call llvm.aarch64.neon.ld2(%ptr) 11324 /// %vec0 = extractelement { <4 x i32>, <4 x i32> } %ld2, i32 0 11325 /// %vec1 = extractelement { <4 x i32>, <4 x i32> } %ld2, i32 1 11326 bool AArch64TargetLowering::lowerInterleavedLoad( 11327 LoadInst *LI, ArrayRef<ShuffleVectorInst *> Shuffles, 11328 ArrayRef<unsigned> Indices, unsigned Factor) const { 11329 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() && 11330 "Invalid interleave factor"); 11331 assert(!Shuffles.empty() && "Empty shufflevector input"); 11332 assert(Shuffles.size() == Indices.size() && 11333 "Unmatched number of shufflevectors and indices"); 11334 11335 const DataLayout &DL = LI->getModule()->getDataLayout(); 11336 11337 VectorType *VTy = Shuffles[0]->getType(); 11338 11339 // Skip if we do not have NEON and skip illegal vector types. We can 11340 // "legalize" wide vector types into multiple interleaved accesses as long as 11341 // the vector types are divisible by 128. 11342 if (!Subtarget->hasNEON() || !isLegalInterleavedAccessType(VTy, DL)) 11343 return false; 11344 11345 unsigned NumLoads = getNumInterleavedAccesses(VTy, DL); 11346 11347 auto *FVTy = cast<FixedVectorType>(VTy); 11348 11349 // A pointer vector can not be the return type of the ldN intrinsics. Need to 11350 // load integer vectors first and then convert to pointer vectors. 11351 Type *EltTy = FVTy->getElementType(); 11352 if (EltTy->isPointerTy()) 11353 FVTy = 11354 FixedVectorType::get(DL.getIntPtrType(EltTy), FVTy->getNumElements()); 11355 11356 IRBuilder<> Builder(LI); 11357 11358 // The base address of the load. 11359 Value *BaseAddr = LI->getPointerOperand(); 11360 11361 if (NumLoads > 1) { 11362 // If we're going to generate more than one load, reset the sub-vector type 11363 // to something legal. 11364 FVTy = FixedVectorType::get(FVTy->getElementType(), 11365 FVTy->getNumElements() / NumLoads); 11366 11367 // We will compute the pointer operand of each load from the original base 11368 // address using GEPs. Cast the base address to a pointer to the scalar 11369 // element type. 11370 BaseAddr = Builder.CreateBitCast( 11371 BaseAddr, 11372 FVTy->getElementType()->getPointerTo(LI->getPointerAddressSpace())); 11373 } 11374 11375 Type *PtrTy = FVTy->getPointerTo(LI->getPointerAddressSpace()); 11376 Type *Tys[2] = {FVTy, PtrTy}; 11377 static const Intrinsic::ID LoadInts[3] = {Intrinsic::aarch64_neon_ld2, 11378 Intrinsic::aarch64_neon_ld3, 11379 Intrinsic::aarch64_neon_ld4}; 11380 Function *LdNFunc = 11381 Intrinsic::getDeclaration(LI->getModule(), LoadInts[Factor - 2], Tys); 11382 11383 // Holds sub-vectors extracted from the load intrinsic return values. The 11384 // sub-vectors are associated with the shufflevector instructions they will 11385 // replace. 11386 DenseMap<ShuffleVectorInst *, SmallVector<Value *, 4>> SubVecs; 11387 11388 for (unsigned LoadCount = 0; LoadCount < NumLoads; ++LoadCount) { 11389 11390 // If we're generating more than one load, compute the base address of 11391 // subsequent loads as an offset from the previous. 11392 if (LoadCount > 0) 11393 BaseAddr = Builder.CreateConstGEP1_32(FVTy->getElementType(), BaseAddr, 11394 FVTy->getNumElements() * Factor); 11395 11396 CallInst *LdN = Builder.CreateCall( 11397 LdNFunc, Builder.CreateBitCast(BaseAddr, PtrTy), "ldN"); 11398 11399 // Extract and store the sub-vectors returned by the load intrinsic. 11400 for (unsigned i = 0; i < Shuffles.size(); i++) { 11401 ShuffleVectorInst *SVI = Shuffles[i]; 11402 unsigned Index = Indices[i]; 11403 11404 Value *SubVec = Builder.CreateExtractValue(LdN, Index); 11405 11406 // Convert the integer vector to pointer vector if the element is pointer. 11407 if (EltTy->isPointerTy()) 11408 SubVec = Builder.CreateIntToPtr( 11409 SubVec, FixedVectorType::get(SVI->getType()->getElementType(), 11410 FVTy->getNumElements())); 11411 SubVecs[SVI].push_back(SubVec); 11412 } 11413 } 11414 11415 // Replace uses of the shufflevector instructions with the sub-vectors 11416 // returned by the load intrinsic. If a shufflevector instruction is 11417 // associated with more than one sub-vector, those sub-vectors will be 11418 // concatenated into a single wide vector. 11419 for (ShuffleVectorInst *SVI : Shuffles) { 11420 auto &SubVec = SubVecs[SVI]; 11421 auto *WideVec = 11422 SubVec.size() > 1 ? concatenateVectors(Builder, SubVec) : SubVec[0]; 11423 SVI->replaceAllUsesWith(WideVec); 11424 } 11425 11426 return true; 11427 } 11428 11429 /// Lower an interleaved store into a stN intrinsic. 11430 /// 11431 /// E.g. Lower an interleaved store (Factor = 3): 11432 /// %i.vec = shuffle <8 x i32> %v0, <8 x i32> %v1, 11433 /// <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> 11434 /// store <12 x i32> %i.vec, <12 x i32>* %ptr 11435 /// 11436 /// Into: 11437 /// %sub.v0 = shuffle <8 x i32> %v0, <8 x i32> v1, <0, 1, 2, 3> 11438 /// %sub.v1 = shuffle <8 x i32> %v0, <8 x i32> v1, <4, 5, 6, 7> 11439 /// %sub.v2 = shuffle <8 x i32> %v0, <8 x i32> v1, <8, 9, 10, 11> 11440 /// call void llvm.aarch64.neon.st3(%sub.v0, %sub.v1, %sub.v2, %ptr) 11441 /// 11442 /// Note that the new shufflevectors will be removed and we'll only generate one 11443 /// st3 instruction in CodeGen. 11444 /// 11445 /// Example for a more general valid mask (Factor 3). Lower: 11446 /// %i.vec = shuffle <32 x i32> %v0, <32 x i32> %v1, 11447 /// <4, 32, 16, 5, 33, 17, 6, 34, 18, 7, 35, 19> 11448 /// store <12 x i32> %i.vec, <12 x i32>* %ptr 11449 /// 11450 /// Into: 11451 /// %sub.v0 = shuffle <32 x i32> %v0, <32 x i32> v1, <4, 5, 6, 7> 11452 /// %sub.v1 = shuffle <32 x i32> %v0, <32 x i32> v1, <32, 33, 34, 35> 11453 /// %sub.v2 = shuffle <32 x i32> %v0, <32 x i32> v1, <16, 17, 18, 19> 11454 /// call void llvm.aarch64.neon.st3(%sub.v0, %sub.v1, %sub.v2, %ptr) 11455 bool AArch64TargetLowering::lowerInterleavedStore(StoreInst *SI, 11456 ShuffleVectorInst *SVI, 11457 unsigned Factor) const { 11458 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() && 11459 "Invalid interleave factor"); 11460 11461 auto *VecTy = cast<FixedVectorType>(SVI->getType()); 11462 assert(VecTy->getNumElements() % Factor == 0 && "Invalid interleaved store"); 11463 11464 unsigned LaneLen = VecTy->getNumElements() / Factor; 11465 Type *EltTy = VecTy->getElementType(); 11466 auto *SubVecTy = FixedVectorType::get(EltTy, LaneLen); 11467 11468 const DataLayout &DL = SI->getModule()->getDataLayout(); 11469 11470 // Skip if we do not have NEON and skip illegal vector types. We can 11471 // "legalize" wide vector types into multiple interleaved accesses as long as 11472 // the vector types are divisible by 128. 11473 if (!Subtarget->hasNEON() || !isLegalInterleavedAccessType(SubVecTy, DL)) 11474 return false; 11475 11476 unsigned NumStores = getNumInterleavedAccesses(SubVecTy, DL); 11477 11478 Value *Op0 = SVI->getOperand(0); 11479 Value *Op1 = SVI->getOperand(1); 11480 IRBuilder<> Builder(SI); 11481 11482 // StN intrinsics don't support pointer vectors as arguments. Convert pointer 11483 // vectors to integer vectors. 11484 if (EltTy->isPointerTy()) { 11485 Type *IntTy = DL.getIntPtrType(EltTy); 11486 unsigned NumOpElts = 11487 cast<FixedVectorType>(Op0->getType())->getNumElements(); 11488 11489 // Convert to the corresponding integer vector. 11490 auto *IntVecTy = FixedVectorType::get(IntTy, NumOpElts); 11491 Op0 = Builder.CreatePtrToInt(Op0, IntVecTy); 11492 Op1 = Builder.CreatePtrToInt(Op1, IntVecTy); 11493 11494 SubVecTy = FixedVectorType::get(IntTy, LaneLen); 11495 } 11496 11497 // The base address of the store. 11498 Value *BaseAddr = SI->getPointerOperand(); 11499 11500 if (NumStores > 1) { 11501 // If we're going to generate more than one store, reset the lane length 11502 // and sub-vector type to something legal. 11503 LaneLen /= NumStores; 11504 SubVecTy = FixedVectorType::get(SubVecTy->getElementType(), LaneLen); 11505 11506 // We will compute the pointer operand of each store from the original base 11507 // address using GEPs. Cast the base address to a pointer to the scalar 11508 // element type. 11509 BaseAddr = Builder.CreateBitCast( 11510 BaseAddr, 11511 SubVecTy->getElementType()->getPointerTo(SI->getPointerAddressSpace())); 11512 } 11513 11514 auto Mask = SVI->getShuffleMask(); 11515 11516 Type *PtrTy = SubVecTy->getPointerTo(SI->getPointerAddressSpace()); 11517 Type *Tys[2] = {SubVecTy, PtrTy}; 11518 static const Intrinsic::ID StoreInts[3] = {Intrinsic::aarch64_neon_st2, 11519 Intrinsic::aarch64_neon_st3, 11520 Intrinsic::aarch64_neon_st4}; 11521 Function *StNFunc = 11522 Intrinsic::getDeclaration(SI->getModule(), StoreInts[Factor - 2], Tys); 11523 11524 for (unsigned StoreCount = 0; StoreCount < NumStores; ++StoreCount) { 11525 11526 SmallVector<Value *, 5> Ops; 11527 11528 // Split the shufflevector operands into sub vectors for the new stN call. 11529 for (unsigned i = 0; i < Factor; i++) { 11530 unsigned IdxI = StoreCount * LaneLen * Factor + i; 11531 if (Mask[IdxI] >= 0) { 11532 Ops.push_back(Builder.CreateShuffleVector( 11533 Op0, Op1, createSequentialMask(Mask[IdxI], LaneLen, 0))); 11534 } else { 11535 unsigned StartMask = 0; 11536 for (unsigned j = 1; j < LaneLen; j++) { 11537 unsigned IdxJ = StoreCount * LaneLen * Factor + j; 11538 if (Mask[IdxJ * Factor + IdxI] >= 0) { 11539 StartMask = Mask[IdxJ * Factor + IdxI] - IdxJ; 11540 break; 11541 } 11542 } 11543 // Note: Filling undef gaps with random elements is ok, since 11544 // those elements were being written anyway (with undefs). 11545 // In the case of all undefs we're defaulting to using elems from 0 11546 // Note: StartMask cannot be negative, it's checked in 11547 // isReInterleaveMask 11548 Ops.push_back(Builder.CreateShuffleVector( 11549 Op0, Op1, createSequentialMask(StartMask, LaneLen, 0))); 11550 } 11551 } 11552 11553 // If we generating more than one store, we compute the base address of 11554 // subsequent stores as an offset from the previous. 11555 if (StoreCount > 0) 11556 BaseAddr = Builder.CreateConstGEP1_32(SubVecTy->getElementType(), 11557 BaseAddr, LaneLen * Factor); 11558 11559 Ops.push_back(Builder.CreateBitCast(BaseAddr, PtrTy)); 11560 Builder.CreateCall(StNFunc, Ops); 11561 } 11562 return true; 11563 } 11564 11565 // Lower an SVE structured load intrinsic returning a tuple type to target 11566 // specific intrinsic taking the same input but returning a multi-result value 11567 // of the split tuple type. 11568 // 11569 // E.g. Lowering an LD3: 11570 // 11571 // call <vscale x 12 x i32> @llvm.aarch64.sve.ld3.nxv12i32( 11572 // <vscale x 4 x i1> %pred, 11573 // <vscale x 4 x i32>* %addr) 11574 // 11575 // Output DAG: 11576 // 11577 // t0: ch = EntryToken 11578 // t2: nxv4i1,ch = CopyFromReg t0, Register:nxv4i1 %0 11579 // t4: i64,ch = CopyFromReg t0, Register:i64 %1 11580 // t5: nxv4i32,nxv4i32,nxv4i32,ch = AArch64ISD::SVE_LD3 t0, t2, t4 11581 // t6: nxv12i32 = concat_vectors t5, t5:1, t5:2 11582 // 11583 // This is called pre-legalization to avoid widening/splitting issues with 11584 // non-power-of-2 tuple types used for LD3, such as nxv12i32. 11585 SDValue AArch64TargetLowering::LowerSVEStructLoad(unsigned Intrinsic, 11586 ArrayRef<SDValue> LoadOps, 11587 EVT VT, SelectionDAG &DAG, 11588 const SDLoc &DL) const { 11589 assert(VT.isScalableVector() && "Can only lower scalable vectors"); 11590 11591 unsigned N, Opcode; 11592 static std::map<unsigned, std::pair<unsigned, unsigned>> IntrinsicMap = { 11593 {Intrinsic::aarch64_sve_ld2, {2, AArch64ISD::SVE_LD2_MERGE_ZERO}}, 11594 {Intrinsic::aarch64_sve_ld3, {3, AArch64ISD::SVE_LD3_MERGE_ZERO}}, 11595 {Intrinsic::aarch64_sve_ld4, {4, AArch64ISD::SVE_LD4_MERGE_ZERO}}}; 11596 11597 std::tie(N, Opcode) = IntrinsicMap[Intrinsic]; 11598 assert(VT.getVectorElementCount().getKnownMinValue() % N == 0 && 11599 "invalid tuple vector type!"); 11600 11601 EVT SplitVT = 11602 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), 11603 VT.getVectorElementCount().divideCoefficientBy(N)); 11604 assert(isTypeLegal(SplitVT)); 11605 11606 SmallVector<EVT, 5> VTs(N, SplitVT); 11607 VTs.push_back(MVT::Other); // Chain 11608 SDVTList NodeTys = DAG.getVTList(VTs); 11609 11610 SDValue PseudoLoad = DAG.getNode(Opcode, DL, NodeTys, LoadOps); 11611 SmallVector<SDValue, 4> PseudoLoadOps; 11612 for (unsigned I = 0; I < N; ++I) 11613 PseudoLoadOps.push_back(SDValue(PseudoLoad.getNode(), I)); 11614 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, PseudoLoadOps); 11615 } 11616 11617 EVT AArch64TargetLowering::getOptimalMemOpType( 11618 const MemOp &Op, const AttributeList &FuncAttributes) const { 11619 bool CanImplicitFloat = 11620 !FuncAttributes.hasFnAttribute(Attribute::NoImplicitFloat); 11621 bool CanUseNEON = Subtarget->hasNEON() && CanImplicitFloat; 11622 bool CanUseFP = Subtarget->hasFPARMv8() && CanImplicitFloat; 11623 // Only use AdvSIMD to implement memset of 32-byte and above. It would have 11624 // taken one instruction to materialize the v2i64 zero and one store (with 11625 // restrictive addressing mode). Just do i64 stores. 11626 bool IsSmallMemset = Op.isMemset() && Op.size() < 32; 11627 auto AlignmentIsAcceptable = [&](EVT VT, Align AlignCheck) { 11628 if (Op.isAligned(AlignCheck)) 11629 return true; 11630 bool Fast; 11631 return allowsMisalignedMemoryAccesses(VT, 0, Align(1), 11632 MachineMemOperand::MONone, &Fast) && 11633 Fast; 11634 }; 11635 11636 if (CanUseNEON && Op.isMemset() && !IsSmallMemset && 11637 AlignmentIsAcceptable(MVT::v2i64, Align(16))) 11638 return MVT::v2i64; 11639 if (CanUseFP && !IsSmallMemset && AlignmentIsAcceptable(MVT::f128, Align(16))) 11640 return MVT::f128; 11641 if (Op.size() >= 8 && AlignmentIsAcceptable(MVT::i64, Align(8))) 11642 return MVT::i64; 11643 if (Op.size() >= 4 && AlignmentIsAcceptable(MVT::i32, Align(4))) 11644 return MVT::i32; 11645 return MVT::Other; 11646 } 11647 11648 LLT AArch64TargetLowering::getOptimalMemOpLLT( 11649 const MemOp &Op, const AttributeList &FuncAttributes) const { 11650 bool CanImplicitFloat = 11651 !FuncAttributes.hasFnAttribute(Attribute::NoImplicitFloat); 11652 bool CanUseNEON = Subtarget->hasNEON() && CanImplicitFloat; 11653 bool CanUseFP = Subtarget->hasFPARMv8() && CanImplicitFloat; 11654 // Only use AdvSIMD to implement memset of 32-byte and above. It would have 11655 // taken one instruction to materialize the v2i64 zero and one store (with 11656 // restrictive addressing mode). Just do i64 stores. 11657 bool IsSmallMemset = Op.isMemset() && Op.size() < 32; 11658 auto AlignmentIsAcceptable = [&](EVT VT, Align AlignCheck) { 11659 if (Op.isAligned(AlignCheck)) 11660 return true; 11661 bool Fast; 11662 return allowsMisalignedMemoryAccesses(VT, 0, Align(1), 11663 MachineMemOperand::MONone, &Fast) && 11664 Fast; 11665 }; 11666 11667 if (CanUseNEON && Op.isMemset() && !IsSmallMemset && 11668 AlignmentIsAcceptable(MVT::v2i64, Align(16))) 11669 return LLT::vector(2, 64); 11670 if (CanUseFP && !IsSmallMemset && AlignmentIsAcceptable(MVT::f128, Align(16))) 11671 return LLT::scalar(128); 11672 if (Op.size() >= 8 && AlignmentIsAcceptable(MVT::i64, Align(8))) 11673 return LLT::scalar(64); 11674 if (Op.size() >= 4 && AlignmentIsAcceptable(MVT::i32, Align(4))) 11675 return LLT::scalar(32); 11676 return LLT(); 11677 } 11678 11679 // 12-bit optionally shifted immediates are legal for adds. 11680 bool AArch64TargetLowering::isLegalAddImmediate(int64_t Immed) const { 11681 if (Immed == std::numeric_limits<int64_t>::min()) { 11682 LLVM_DEBUG(dbgs() << "Illegal add imm " << Immed 11683 << ": avoid UB for INT64_MIN\n"); 11684 return false; 11685 } 11686 // Same encoding for add/sub, just flip the sign. 11687 Immed = std::abs(Immed); 11688 bool IsLegal = ((Immed >> 12) == 0 || 11689 ((Immed & 0xfff) == 0 && Immed >> 24 == 0)); 11690 LLVM_DEBUG(dbgs() << "Is " << Immed 11691 << " legal add imm: " << (IsLegal ? "yes" : "no") << "\n"); 11692 return IsLegal; 11693 } 11694 11695 // Integer comparisons are implemented with ADDS/SUBS, so the range of valid 11696 // immediates is the same as for an add or a sub. 11697 bool AArch64TargetLowering::isLegalICmpImmediate(int64_t Immed) const { 11698 return isLegalAddImmediate(Immed); 11699 } 11700 11701 /// isLegalAddressingMode - Return true if the addressing mode represented 11702 /// by AM is legal for this target, for a load/store of the specified type. 11703 bool AArch64TargetLowering::isLegalAddressingMode(const DataLayout &DL, 11704 const AddrMode &AM, Type *Ty, 11705 unsigned AS, Instruction *I) const { 11706 // AArch64 has five basic addressing modes: 11707 // reg 11708 // reg + 9-bit signed offset 11709 // reg + SIZE_IN_BYTES * 12-bit unsigned offset 11710 // reg1 + reg2 11711 // reg + SIZE_IN_BYTES * reg 11712 11713 // No global is ever allowed as a base. 11714 if (AM.BaseGV) 11715 return false; 11716 11717 // No reg+reg+imm addressing. 11718 if (AM.HasBaseReg && AM.BaseOffs && AM.Scale) 11719 return false; 11720 11721 // FIXME: Update this method to support scalable addressing modes. 11722 if (isa<ScalableVectorType>(Ty)) 11723 return AM.HasBaseReg && !AM.BaseOffs && !AM.Scale; 11724 11725 // check reg + imm case: 11726 // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12 11727 uint64_t NumBytes = 0; 11728 if (Ty->isSized()) { 11729 uint64_t NumBits = DL.getTypeSizeInBits(Ty); 11730 NumBytes = NumBits / 8; 11731 if (!isPowerOf2_64(NumBits)) 11732 NumBytes = 0; 11733 } 11734 11735 if (!AM.Scale) { 11736 int64_t Offset = AM.BaseOffs; 11737 11738 // 9-bit signed offset 11739 if (isInt<9>(Offset)) 11740 return true; 11741 11742 // 12-bit unsigned offset 11743 unsigned shift = Log2_64(NumBytes); 11744 if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 && 11745 // Must be a multiple of NumBytes (NumBytes is a power of 2) 11746 (Offset >> shift) << shift == Offset) 11747 return true; 11748 return false; 11749 } 11750 11751 // Check reg1 + SIZE_IN_BYTES * reg2 and reg1 + reg2 11752 11753 return AM.Scale == 1 || (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes); 11754 } 11755 11756 bool AArch64TargetLowering::shouldConsiderGEPOffsetSplit() const { 11757 // Consider splitting large offset of struct or array. 11758 return true; 11759 } 11760 11761 InstructionCost AArch64TargetLowering::getScalingFactorCost( 11762 const DataLayout &DL, const AddrMode &AM, Type *Ty, unsigned AS) const { 11763 // Scaling factors are not free at all. 11764 // Operands | Rt Latency 11765 // ------------------------------------------- 11766 // Rt, [Xn, Xm] | 4 11767 // ------------------------------------------- 11768 // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5 11769 // Rt, [Xn, Wm, <extend> #imm] | 11770 if (isLegalAddressingMode(DL, AM, Ty, AS)) 11771 // Scale represents reg2 * scale, thus account for 1 if 11772 // it is not equal to 0 or 1. 11773 return AM.Scale != 0 && AM.Scale != 1; 11774 return -1; 11775 } 11776 11777 bool AArch64TargetLowering::isFMAFasterThanFMulAndFAdd( 11778 const MachineFunction &MF, EVT VT) const { 11779 VT = VT.getScalarType(); 11780 11781 if (!VT.isSimple()) 11782 return false; 11783 11784 switch (VT.getSimpleVT().SimpleTy) { 11785 case MVT::f16: 11786 return Subtarget->hasFullFP16(); 11787 case MVT::f32: 11788 case MVT::f64: 11789 return true; 11790 default: 11791 break; 11792 } 11793 11794 return false; 11795 } 11796 11797 bool AArch64TargetLowering::isFMAFasterThanFMulAndFAdd(const Function &F, 11798 Type *Ty) const { 11799 switch (Ty->getScalarType()->getTypeID()) { 11800 case Type::FloatTyID: 11801 case Type::DoubleTyID: 11802 return true; 11803 default: 11804 return false; 11805 } 11806 } 11807 11808 bool AArch64TargetLowering::generateFMAsInMachineCombiner( 11809 EVT VT, CodeGenOpt::Level OptLevel) const { 11810 return (OptLevel >= CodeGenOpt::Aggressive) && !VT.isScalableVector(); 11811 } 11812 11813 const MCPhysReg * 11814 AArch64TargetLowering::getScratchRegisters(CallingConv::ID) const { 11815 // LR is a callee-save register, but we must treat it as clobbered by any call 11816 // site. Hence we include LR in the scratch registers, which are in turn added 11817 // as implicit-defs for stackmaps and patchpoints. 11818 static const MCPhysReg ScratchRegs[] = { 11819 AArch64::X16, AArch64::X17, AArch64::LR, 0 11820 }; 11821 return ScratchRegs; 11822 } 11823 11824 bool 11825 AArch64TargetLowering::isDesirableToCommuteWithShift(const SDNode *N, 11826 CombineLevel Level) const { 11827 N = N->getOperand(0).getNode(); 11828 EVT VT = N->getValueType(0); 11829 // If N is unsigned bit extraction: ((x >> C) & mask), then do not combine 11830 // it with shift to let it be lowered to UBFX. 11831 if (N->getOpcode() == ISD::AND && (VT == MVT::i32 || VT == MVT::i64) && 11832 isa<ConstantSDNode>(N->getOperand(1))) { 11833 uint64_t TruncMask = N->getConstantOperandVal(1); 11834 if (isMask_64(TruncMask) && 11835 N->getOperand(0).getOpcode() == ISD::SRL && 11836 isa<ConstantSDNode>(N->getOperand(0)->getOperand(1))) 11837 return false; 11838 } 11839 return true; 11840 } 11841 11842 bool AArch64TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm, 11843 Type *Ty) const { 11844 assert(Ty->isIntegerTy()); 11845 11846 unsigned BitSize = Ty->getPrimitiveSizeInBits(); 11847 if (BitSize == 0) 11848 return false; 11849 11850 int64_t Val = Imm.getSExtValue(); 11851 if (Val == 0 || AArch64_AM::isLogicalImmediate(Val, BitSize)) 11852 return true; 11853 11854 if ((int64_t)Val < 0) 11855 Val = ~Val; 11856 if (BitSize == 32) 11857 Val &= (1LL << 32) - 1; 11858 11859 unsigned LZ = countLeadingZeros((uint64_t)Val); 11860 unsigned Shift = (63 - LZ) / 16; 11861 // MOVZ is free so return true for one or fewer MOVK. 11862 return Shift < 3; 11863 } 11864 11865 bool AArch64TargetLowering::isExtractSubvectorCheap(EVT ResVT, EVT SrcVT, 11866 unsigned Index) const { 11867 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT)) 11868 return false; 11869 11870 return (Index == 0 || Index == ResVT.getVectorNumElements()); 11871 } 11872 11873 /// Turn vector tests of the signbit in the form of: 11874 /// xor (sra X, elt_size(X)-1), -1 11875 /// into: 11876 /// cmge X, X, #0 11877 static SDValue foldVectorXorShiftIntoCmp(SDNode *N, SelectionDAG &DAG, 11878 const AArch64Subtarget *Subtarget) { 11879 EVT VT = N->getValueType(0); 11880 if (!Subtarget->hasNEON() || !VT.isVector()) 11881 return SDValue(); 11882 11883 // There must be a shift right algebraic before the xor, and the xor must be a 11884 // 'not' operation. 11885 SDValue Shift = N->getOperand(0); 11886 SDValue Ones = N->getOperand(1); 11887 if (Shift.getOpcode() != AArch64ISD::VASHR || !Shift.hasOneUse() || 11888 !ISD::isBuildVectorAllOnes(Ones.getNode())) 11889 return SDValue(); 11890 11891 // The shift should be smearing the sign bit across each vector element. 11892 auto *ShiftAmt = dyn_cast<ConstantSDNode>(Shift.getOperand(1)); 11893 EVT ShiftEltTy = Shift.getValueType().getVectorElementType(); 11894 if (!ShiftAmt || ShiftAmt->getZExtValue() != ShiftEltTy.getSizeInBits() - 1) 11895 return SDValue(); 11896 11897 return DAG.getNode(AArch64ISD::CMGEz, SDLoc(N), VT, Shift.getOperand(0)); 11898 } 11899 11900 // Turn a v8i8/v16i8 extended vecreduce into a udot/sdot and vecreduce 11901 // vecreduce.add(ext(A)) to vecreduce.add(DOT(zero, A, one)) 11902 // vecreduce.add(mul(ext(A), ext(B))) to vecreduce.add(DOT(zero, A, B)) 11903 static SDValue performVecReduceAddCombine(SDNode *N, SelectionDAG &DAG, 11904 const AArch64Subtarget *ST) { 11905 SDValue Op0 = N->getOperand(0); 11906 if (!ST->hasDotProd() || N->getValueType(0) != MVT::i32 || 11907 Op0.getValueType().getVectorElementType() != MVT::i32) 11908 return SDValue(); 11909 11910 unsigned ExtOpcode = Op0.getOpcode(); 11911 SDValue A = Op0; 11912 SDValue B; 11913 if (ExtOpcode == ISD::MUL) { 11914 A = Op0.getOperand(0); 11915 B = Op0.getOperand(1); 11916 if (A.getOpcode() != B.getOpcode() || 11917 A.getOperand(0).getValueType() != B.getOperand(0).getValueType()) 11918 return SDValue(); 11919 ExtOpcode = A.getOpcode(); 11920 } 11921 if (ExtOpcode != ISD::ZERO_EXTEND && ExtOpcode != ISD::SIGN_EXTEND) 11922 return SDValue(); 11923 11924 EVT Op0VT = A.getOperand(0).getValueType(); 11925 if (Op0VT != MVT::v8i8 && Op0VT != MVT::v16i8) 11926 return SDValue(); 11927 11928 SDLoc DL(Op0); 11929 // For non-mla reductions B can be set to 1. For MLA we take the operand of 11930 // the extend B. 11931 if (!B) 11932 B = DAG.getConstant(1, DL, Op0VT); 11933 else 11934 B = B.getOperand(0); 11935 11936 SDValue Zeros = 11937 DAG.getConstant(0, DL, Op0VT == MVT::v8i8 ? MVT::v2i32 : MVT::v4i32); 11938 auto DotOpcode = 11939 (ExtOpcode == ISD::ZERO_EXTEND) ? AArch64ISD::UDOT : AArch64ISD::SDOT; 11940 SDValue Dot = DAG.getNode(DotOpcode, DL, Zeros.getValueType(), Zeros, 11941 A.getOperand(0), B); 11942 return DAG.getNode(ISD::VECREDUCE_ADD, DL, N->getValueType(0), Dot); 11943 } 11944 11945 // Given a ABS node, detect the following pattern: 11946 // (ABS (SUB (EXTEND a), (EXTEND b))). 11947 // Generates UABD/SABD instruction. 11948 static SDValue performABSCombine(SDNode *N, SelectionDAG &DAG, 11949 TargetLowering::DAGCombinerInfo &DCI, 11950 const AArch64Subtarget *Subtarget) { 11951 SDValue AbsOp1 = N->getOperand(0); 11952 SDValue Op0, Op1; 11953 11954 if (AbsOp1.getOpcode() != ISD::SUB) 11955 return SDValue(); 11956 11957 Op0 = AbsOp1.getOperand(0); 11958 Op1 = AbsOp1.getOperand(1); 11959 11960 unsigned Opc0 = Op0.getOpcode(); 11961 // Check if the operands of the sub are (zero|sign)-extended. 11962 if (Opc0 != Op1.getOpcode() || 11963 (Opc0 != ISD::ZERO_EXTEND && Opc0 != ISD::SIGN_EXTEND)) 11964 return SDValue(); 11965 11966 EVT VectorT1 = Op0.getOperand(0).getValueType(); 11967 EVT VectorT2 = Op1.getOperand(0).getValueType(); 11968 // Check if vectors are of same type and valid size. 11969 uint64_t Size = VectorT1.getFixedSizeInBits(); 11970 if (VectorT1 != VectorT2 || (Size != 64 && Size != 128)) 11971 return SDValue(); 11972 11973 // Check if vector element types are valid. 11974 EVT VT1 = VectorT1.getVectorElementType(); 11975 if (VT1 != MVT::i8 && VT1 != MVT::i16 && VT1 != MVT::i32) 11976 return SDValue(); 11977 11978 Op0 = Op0.getOperand(0); 11979 Op1 = Op1.getOperand(0); 11980 unsigned ABDOpcode = 11981 (Opc0 == ISD::SIGN_EXTEND) ? AArch64ISD::SABD : AArch64ISD::UABD; 11982 SDValue ABD = 11983 DAG.getNode(ABDOpcode, SDLoc(N), Op0->getValueType(0), Op0, Op1); 11984 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0), ABD); 11985 } 11986 11987 static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG, 11988 TargetLowering::DAGCombinerInfo &DCI, 11989 const AArch64Subtarget *Subtarget) { 11990 if (DCI.isBeforeLegalizeOps()) 11991 return SDValue(); 11992 11993 return foldVectorXorShiftIntoCmp(N, DAG, Subtarget); 11994 } 11995 11996 SDValue 11997 AArch64TargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor, 11998 SelectionDAG &DAG, 11999 SmallVectorImpl<SDNode *> &Created) const { 12000 AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); 12001 if (isIntDivCheap(N->getValueType(0), Attr)) 12002 return SDValue(N,0); // Lower SDIV as SDIV 12003 12004 // fold (sdiv X, pow2) 12005 EVT VT = N->getValueType(0); 12006 if ((VT != MVT::i32 && VT != MVT::i64) || 12007 !(Divisor.isPowerOf2() || (-Divisor).isPowerOf2())) 12008 return SDValue(); 12009 12010 SDLoc DL(N); 12011 SDValue N0 = N->getOperand(0); 12012 unsigned Lg2 = Divisor.countTrailingZeros(); 12013 SDValue Zero = DAG.getConstant(0, DL, VT); 12014 SDValue Pow2MinusOne = DAG.getConstant((1ULL << Lg2) - 1, DL, VT); 12015 12016 // Add (N0 < 0) ? Pow2 - 1 : 0; 12017 SDValue CCVal; 12018 SDValue Cmp = getAArch64Cmp(N0, Zero, ISD::SETLT, CCVal, DAG, DL); 12019 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0, Pow2MinusOne); 12020 SDValue CSel = DAG.getNode(AArch64ISD::CSEL, DL, VT, Add, N0, CCVal, Cmp); 12021 12022 Created.push_back(Cmp.getNode()); 12023 Created.push_back(Add.getNode()); 12024 Created.push_back(CSel.getNode()); 12025 12026 // Divide by pow2. 12027 SDValue SRA = 12028 DAG.getNode(ISD::SRA, DL, VT, CSel, DAG.getConstant(Lg2, DL, MVT::i64)); 12029 12030 // If we're dividing by a positive value, we're done. Otherwise, we must 12031 // negate the result. 12032 if (Divisor.isNonNegative()) 12033 return SRA; 12034 12035 Created.push_back(SRA.getNode()); 12036 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), SRA); 12037 } 12038 12039 static bool IsSVECntIntrinsic(SDValue S) { 12040 switch(getIntrinsicID(S.getNode())) { 12041 default: 12042 break; 12043 case Intrinsic::aarch64_sve_cntb: 12044 case Intrinsic::aarch64_sve_cnth: 12045 case Intrinsic::aarch64_sve_cntw: 12046 case Intrinsic::aarch64_sve_cntd: 12047 return true; 12048 } 12049 return false; 12050 } 12051 12052 /// Calculates what the pre-extend type is, based on the extension 12053 /// operation node provided by \p Extend. 12054 /// 12055 /// In the case that \p Extend is a SIGN_EXTEND or a ZERO_EXTEND, the 12056 /// pre-extend type is pulled directly from the operand, while other extend 12057 /// operations need a bit more inspection to get this information. 12058 /// 12059 /// \param Extend The SDNode from the DAG that represents the extend operation 12060 /// \param DAG The SelectionDAG hosting the \p Extend node 12061 /// 12062 /// \returns The type representing the \p Extend source type, or \p MVT::Other 12063 /// if no valid type can be determined 12064 static EVT calculatePreExtendType(SDValue Extend, SelectionDAG &DAG) { 12065 switch (Extend.getOpcode()) { 12066 case ISD::SIGN_EXTEND: 12067 case ISD::ZERO_EXTEND: 12068 return Extend.getOperand(0).getValueType(); 12069 case ISD::AssertSext: 12070 case ISD::AssertZext: 12071 case ISD::SIGN_EXTEND_INREG: { 12072 VTSDNode *TypeNode = dyn_cast<VTSDNode>(Extend.getOperand(1)); 12073 if (!TypeNode) 12074 return MVT::Other; 12075 return TypeNode->getVT(); 12076 } 12077 case ISD::AND: { 12078 ConstantSDNode *Constant = 12079 dyn_cast<ConstantSDNode>(Extend.getOperand(1).getNode()); 12080 if (!Constant) 12081 return MVT::Other; 12082 12083 uint32_t Mask = Constant->getZExtValue(); 12084 12085 if (Mask == UCHAR_MAX) 12086 return MVT::i8; 12087 else if (Mask == USHRT_MAX) 12088 return MVT::i16; 12089 else if (Mask == UINT_MAX) 12090 return MVT::i32; 12091 12092 return MVT::Other; 12093 } 12094 default: 12095 return MVT::Other; 12096 } 12097 12098 llvm_unreachable("Code path unhandled in calculatePreExtendType!"); 12099 } 12100 12101 /// Combines a dup(sext/zext) node pattern into sext/zext(dup) 12102 /// making use of the vector SExt/ZExt rather than the scalar SExt/ZExt 12103 static SDValue performCommonVectorExtendCombine(SDValue VectorShuffle, 12104 SelectionDAG &DAG) { 12105 12106 ShuffleVectorSDNode *ShuffleNode = 12107 dyn_cast<ShuffleVectorSDNode>(VectorShuffle.getNode()); 12108 if (!ShuffleNode) 12109 return SDValue(); 12110 12111 // Ensuring the mask is zero before continuing 12112 if (!ShuffleNode->isSplat() || ShuffleNode->getSplatIndex() != 0) 12113 return SDValue(); 12114 12115 SDValue InsertVectorElt = VectorShuffle.getOperand(0); 12116 12117 if (InsertVectorElt.getOpcode() != ISD::INSERT_VECTOR_ELT) 12118 return SDValue(); 12119 12120 SDValue InsertLane = InsertVectorElt.getOperand(2); 12121 ConstantSDNode *Constant = dyn_cast<ConstantSDNode>(InsertLane.getNode()); 12122 // Ensures the insert is inserting into lane 0 12123 if (!Constant || Constant->getZExtValue() != 0) 12124 return SDValue(); 12125 12126 SDValue Extend = InsertVectorElt.getOperand(1); 12127 unsigned ExtendOpcode = Extend.getOpcode(); 12128 12129 bool IsSExt = ExtendOpcode == ISD::SIGN_EXTEND || 12130 ExtendOpcode == ISD::SIGN_EXTEND_INREG || 12131 ExtendOpcode == ISD::AssertSext; 12132 if (!IsSExt && ExtendOpcode != ISD::ZERO_EXTEND && 12133 ExtendOpcode != ISD::AssertZext && ExtendOpcode != ISD::AND) 12134 return SDValue(); 12135 12136 EVT TargetType = VectorShuffle.getValueType(); 12137 EVT PreExtendType = calculatePreExtendType(Extend, DAG); 12138 12139 if ((TargetType != MVT::v8i16 && TargetType != MVT::v4i32 && 12140 TargetType != MVT::v2i64) || 12141 (PreExtendType == MVT::Other)) 12142 return SDValue(); 12143 12144 // Restrict valid pre-extend data type 12145 if (PreExtendType != MVT::i8 && PreExtendType != MVT::i16 && 12146 PreExtendType != MVT::i32) 12147 return SDValue(); 12148 12149 EVT PreExtendVT = TargetType.changeVectorElementType(PreExtendType); 12150 12151 if (PreExtendVT.getVectorElementCount() != TargetType.getVectorElementCount()) 12152 return SDValue(); 12153 12154 if (TargetType.getScalarSizeInBits() != PreExtendVT.getScalarSizeInBits() * 2) 12155 return SDValue(); 12156 12157 SDLoc DL(VectorShuffle); 12158 12159 SDValue InsertVectorNode = DAG.getNode( 12160 InsertVectorElt.getOpcode(), DL, PreExtendVT, DAG.getUNDEF(PreExtendVT), 12161 DAG.getAnyExtOrTrunc(Extend.getOperand(0), DL, PreExtendType), 12162 DAG.getConstant(0, DL, MVT::i64)); 12163 12164 std::vector<int> ShuffleMask(TargetType.getVectorElementCount().getValue()); 12165 12166 SDValue VectorShuffleNode = 12167 DAG.getVectorShuffle(PreExtendVT, DL, InsertVectorNode, 12168 DAG.getUNDEF(PreExtendVT), ShuffleMask); 12169 12170 SDValue ExtendNode = DAG.getNode(IsSExt ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND, 12171 DL, TargetType, VectorShuffleNode); 12172 12173 return ExtendNode; 12174 } 12175 12176 /// Combines a mul(dup(sext/zext)) node pattern into mul(sext/zext(dup)) 12177 /// making use of the vector SExt/ZExt rather than the scalar SExt/ZExt 12178 static SDValue performMulVectorExtendCombine(SDNode *Mul, SelectionDAG &DAG) { 12179 // If the value type isn't a vector, none of the operands are going to be dups 12180 if (!Mul->getValueType(0).isVector()) 12181 return SDValue(); 12182 12183 SDValue Op0 = performCommonVectorExtendCombine(Mul->getOperand(0), DAG); 12184 SDValue Op1 = performCommonVectorExtendCombine(Mul->getOperand(1), DAG); 12185 12186 // Neither operands have been changed, don't make any further changes 12187 if (!Op0 && !Op1) 12188 return SDValue(); 12189 12190 SDLoc DL(Mul); 12191 return DAG.getNode(Mul->getOpcode(), DL, Mul->getValueType(0), 12192 Op0 ? Op0 : Mul->getOperand(0), 12193 Op1 ? Op1 : Mul->getOperand(1)); 12194 } 12195 12196 static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG, 12197 TargetLowering::DAGCombinerInfo &DCI, 12198 const AArch64Subtarget *Subtarget) { 12199 12200 if (SDValue Ext = performMulVectorExtendCombine(N, DAG)) 12201 return Ext; 12202 12203 if (DCI.isBeforeLegalizeOps()) 12204 return SDValue(); 12205 12206 // The below optimizations require a constant RHS. 12207 if (!isa<ConstantSDNode>(N->getOperand(1))) 12208 return SDValue(); 12209 12210 SDValue N0 = N->getOperand(0); 12211 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(1)); 12212 const APInt &ConstValue = C->getAPIntValue(); 12213 12214 // Allow the scaling to be folded into the `cnt` instruction by preventing 12215 // the scaling to be obscured here. This makes it easier to pattern match. 12216 if (IsSVECntIntrinsic(N0) || 12217 (N0->getOpcode() == ISD::TRUNCATE && 12218 (IsSVECntIntrinsic(N0->getOperand(0))))) 12219 if (ConstValue.sge(1) && ConstValue.sle(16)) 12220 return SDValue(); 12221 12222 // Multiplication of a power of two plus/minus one can be done more 12223 // cheaply as as shift+add/sub. For now, this is true unilaterally. If 12224 // future CPUs have a cheaper MADD instruction, this may need to be 12225 // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and 12226 // 64-bit is 5 cycles, so this is always a win. 12227 // More aggressively, some multiplications N0 * C can be lowered to 12228 // shift+add+shift if the constant C = A * B where A = 2^N + 1 and B = 2^M, 12229 // e.g. 6=3*2=(2+1)*2. 12230 // TODO: consider lowering more cases, e.g. C = 14, -6, -14 or even 45 12231 // which equals to (1+2)*16-(1+2). 12232 // TrailingZeroes is used to test if the mul can be lowered to 12233 // shift+add+shift. 12234 unsigned TrailingZeroes = ConstValue.countTrailingZeros(); 12235 if (TrailingZeroes) { 12236 // Conservatively do not lower to shift+add+shift if the mul might be 12237 // folded into smul or umul. 12238 if (N0->hasOneUse() && (isSignExtended(N0.getNode(), DAG) || 12239 isZeroExtended(N0.getNode(), DAG))) 12240 return SDValue(); 12241 // Conservatively do not lower to shift+add+shift if the mul might be 12242 // folded into madd or msub. 12243 if (N->hasOneUse() && (N->use_begin()->getOpcode() == ISD::ADD || 12244 N->use_begin()->getOpcode() == ISD::SUB)) 12245 return SDValue(); 12246 } 12247 // Use ShiftedConstValue instead of ConstValue to support both shift+add/sub 12248 // and shift+add+shift. 12249 APInt ShiftedConstValue = ConstValue.ashr(TrailingZeroes); 12250 12251 unsigned ShiftAmt, AddSubOpc; 12252 // Is the shifted value the LHS operand of the add/sub? 12253 bool ShiftValUseIsN0 = true; 12254 // Do we need to negate the result? 12255 bool NegateResult = false; 12256 12257 if (ConstValue.isNonNegative()) { 12258 // (mul x, 2^N + 1) => (add (shl x, N), x) 12259 // (mul x, 2^N - 1) => (sub (shl x, N), x) 12260 // (mul x, (2^N + 1) * 2^M) => (shl (add (shl x, N), x), M) 12261 APInt SCVMinus1 = ShiftedConstValue - 1; 12262 APInt CVPlus1 = ConstValue + 1; 12263 if (SCVMinus1.isPowerOf2()) { 12264 ShiftAmt = SCVMinus1.logBase2(); 12265 AddSubOpc = ISD::ADD; 12266 } else if (CVPlus1.isPowerOf2()) { 12267 ShiftAmt = CVPlus1.logBase2(); 12268 AddSubOpc = ISD::SUB; 12269 } else 12270 return SDValue(); 12271 } else { 12272 // (mul x, -(2^N - 1)) => (sub x, (shl x, N)) 12273 // (mul x, -(2^N + 1)) => - (add (shl x, N), x) 12274 APInt CVNegPlus1 = -ConstValue + 1; 12275 APInt CVNegMinus1 = -ConstValue - 1; 12276 if (CVNegPlus1.isPowerOf2()) { 12277 ShiftAmt = CVNegPlus1.logBase2(); 12278 AddSubOpc = ISD::SUB; 12279 ShiftValUseIsN0 = false; 12280 } else if (CVNegMinus1.isPowerOf2()) { 12281 ShiftAmt = CVNegMinus1.logBase2(); 12282 AddSubOpc = ISD::ADD; 12283 NegateResult = true; 12284 } else 12285 return SDValue(); 12286 } 12287 12288 SDLoc DL(N); 12289 EVT VT = N->getValueType(0); 12290 SDValue ShiftedVal = DAG.getNode(ISD::SHL, DL, VT, N0, 12291 DAG.getConstant(ShiftAmt, DL, MVT::i64)); 12292 12293 SDValue AddSubN0 = ShiftValUseIsN0 ? ShiftedVal : N0; 12294 SDValue AddSubN1 = ShiftValUseIsN0 ? N0 : ShiftedVal; 12295 SDValue Res = DAG.getNode(AddSubOpc, DL, VT, AddSubN0, AddSubN1); 12296 assert(!(NegateResult && TrailingZeroes) && 12297 "NegateResult and TrailingZeroes cannot both be true for now."); 12298 // Negate the result. 12299 if (NegateResult) 12300 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Res); 12301 // Shift the result. 12302 if (TrailingZeroes) 12303 return DAG.getNode(ISD::SHL, DL, VT, Res, 12304 DAG.getConstant(TrailingZeroes, DL, MVT::i64)); 12305 return Res; 12306 } 12307 12308 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N, 12309 SelectionDAG &DAG) { 12310 // Take advantage of vector comparisons producing 0 or -1 in each lane to 12311 // optimize away operation when it's from a constant. 12312 // 12313 // The general transformation is: 12314 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) --> 12315 // AND(VECTOR_CMP(x,y), constant2) 12316 // constant2 = UNARYOP(constant) 12317 12318 // Early exit if this isn't a vector operation, the operand of the 12319 // unary operation isn't a bitwise AND, or if the sizes of the operations 12320 // aren't the same. 12321 EVT VT = N->getValueType(0); 12322 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND || 12323 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC || 12324 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits()) 12325 return SDValue(); 12326 12327 // Now check that the other operand of the AND is a constant. We could 12328 // make the transformation for non-constant splats as well, but it's unclear 12329 // that would be a benefit as it would not eliminate any operations, just 12330 // perform one more step in scalar code before moving to the vector unit. 12331 if (BuildVectorSDNode *BV = 12332 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) { 12333 // Bail out if the vector isn't a constant. 12334 if (!BV->isConstant()) 12335 return SDValue(); 12336 12337 // Everything checks out. Build up the new and improved node. 12338 SDLoc DL(N); 12339 EVT IntVT = BV->getValueType(0); 12340 // Create a new constant of the appropriate type for the transformed 12341 // DAG. 12342 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0)); 12343 // The AND node needs bitcasts to/from an integer vector type around it. 12344 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst); 12345 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT, 12346 N->getOperand(0)->getOperand(0), MaskConst); 12347 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd); 12348 return Res; 12349 } 12350 12351 return SDValue(); 12352 } 12353 12354 static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG, 12355 const AArch64Subtarget *Subtarget) { 12356 // First try to optimize away the conversion when it's conditionally from 12357 // a constant. Vectors only. 12358 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG)) 12359 return Res; 12360 12361 EVT VT = N->getValueType(0); 12362 if (VT != MVT::f32 && VT != MVT::f64) 12363 return SDValue(); 12364 12365 // Only optimize when the source and destination types have the same width. 12366 if (VT.getSizeInBits() != N->getOperand(0).getValueSizeInBits()) 12367 return SDValue(); 12368 12369 // If the result of an integer load is only used by an integer-to-float 12370 // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead. 12371 // This eliminates an "integer-to-vector-move" UOP and improves throughput. 12372 SDValue N0 = N->getOperand(0); 12373 if (Subtarget->hasNEON() && ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() && 12374 // Do not change the width of a volatile load. 12375 !cast<LoadSDNode>(N0)->isVolatile()) { 12376 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 12377 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(), 12378 LN0->getPointerInfo(), LN0->getAlignment(), 12379 LN0->getMemOperand()->getFlags()); 12380 12381 // Make sure successors of the original load stay after it by updating them 12382 // to use the new Chain. 12383 DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1)); 12384 12385 unsigned Opcode = 12386 (N->getOpcode() == ISD::SINT_TO_FP) ? AArch64ISD::SITOF : AArch64ISD::UITOF; 12387 return DAG.getNode(Opcode, SDLoc(N), VT, Load); 12388 } 12389 12390 return SDValue(); 12391 } 12392 12393 /// Fold a floating-point multiply by power of two into floating-point to 12394 /// fixed-point conversion. 12395 static SDValue performFpToIntCombine(SDNode *N, SelectionDAG &DAG, 12396 TargetLowering::DAGCombinerInfo &DCI, 12397 const AArch64Subtarget *Subtarget) { 12398 if (!Subtarget->hasNEON()) 12399 return SDValue(); 12400 12401 if (!N->getValueType(0).isSimple()) 12402 return SDValue(); 12403 12404 SDValue Op = N->getOperand(0); 12405 if (!Op.getValueType().isVector() || !Op.getValueType().isSimple() || 12406 Op.getOpcode() != ISD::FMUL) 12407 return SDValue(); 12408 12409 SDValue ConstVec = Op->getOperand(1); 12410 if (!isa<BuildVectorSDNode>(ConstVec)) 12411 return SDValue(); 12412 12413 MVT FloatTy = Op.getSimpleValueType().getVectorElementType(); 12414 uint32_t FloatBits = FloatTy.getSizeInBits(); 12415 if (FloatBits != 32 && FloatBits != 64) 12416 return SDValue(); 12417 12418 MVT IntTy = N->getSimpleValueType(0).getVectorElementType(); 12419 uint32_t IntBits = IntTy.getSizeInBits(); 12420 if (IntBits != 16 && IntBits != 32 && IntBits != 64) 12421 return SDValue(); 12422 12423 // Avoid conversions where iN is larger than the float (e.g., float -> i64). 12424 if (IntBits > FloatBits) 12425 return SDValue(); 12426 12427 BitVector UndefElements; 12428 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(ConstVec); 12429 int32_t Bits = IntBits == 64 ? 64 : 32; 12430 int32_t C = BV->getConstantFPSplatPow2ToLog2Int(&UndefElements, Bits + 1); 12431 if (C == -1 || C == 0 || C > Bits) 12432 return SDValue(); 12433 12434 MVT ResTy; 12435 unsigned NumLanes = Op.getValueType().getVectorNumElements(); 12436 switch (NumLanes) { 12437 default: 12438 return SDValue(); 12439 case 2: 12440 ResTy = FloatBits == 32 ? MVT::v2i32 : MVT::v2i64; 12441 break; 12442 case 4: 12443 ResTy = FloatBits == 32 ? MVT::v4i32 : MVT::v4i64; 12444 break; 12445 } 12446 12447 if (ResTy == MVT::v4i64 && DCI.isBeforeLegalizeOps()) 12448 return SDValue(); 12449 12450 assert((ResTy != MVT::v4i64 || DCI.isBeforeLegalizeOps()) && 12451 "Illegal vector type after legalization"); 12452 12453 SDLoc DL(N); 12454 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT; 12455 unsigned IntrinsicOpcode = IsSigned ? Intrinsic::aarch64_neon_vcvtfp2fxs 12456 : Intrinsic::aarch64_neon_vcvtfp2fxu; 12457 SDValue FixConv = 12458 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, ResTy, 12459 DAG.getConstant(IntrinsicOpcode, DL, MVT::i32), 12460 Op->getOperand(0), DAG.getConstant(C, DL, MVT::i32)); 12461 // We can handle smaller integers by generating an extra trunc. 12462 if (IntBits < FloatBits) 12463 FixConv = DAG.getNode(ISD::TRUNCATE, DL, N->getValueType(0), FixConv); 12464 12465 return FixConv; 12466 } 12467 12468 /// Fold a floating-point divide by power of two into fixed-point to 12469 /// floating-point conversion. 12470 static SDValue performFDivCombine(SDNode *N, SelectionDAG &DAG, 12471 TargetLowering::DAGCombinerInfo &DCI, 12472 const AArch64Subtarget *Subtarget) { 12473 if (!Subtarget->hasNEON()) 12474 return SDValue(); 12475 12476 SDValue Op = N->getOperand(0); 12477 unsigned Opc = Op->getOpcode(); 12478 if (!Op.getValueType().isVector() || !Op.getValueType().isSimple() || 12479 !Op.getOperand(0).getValueType().isSimple() || 12480 (Opc != ISD::SINT_TO_FP && Opc != ISD::UINT_TO_FP)) 12481 return SDValue(); 12482 12483 SDValue ConstVec = N->getOperand(1); 12484 if (!isa<BuildVectorSDNode>(ConstVec)) 12485 return SDValue(); 12486 12487 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType(); 12488 int32_t IntBits = IntTy.getSizeInBits(); 12489 if (IntBits != 16 && IntBits != 32 && IntBits != 64) 12490 return SDValue(); 12491 12492 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType(); 12493 int32_t FloatBits = FloatTy.getSizeInBits(); 12494 if (FloatBits != 32 && FloatBits != 64) 12495 return SDValue(); 12496 12497 // Avoid conversions where iN is larger than the float (e.g., i64 -> float). 12498 if (IntBits > FloatBits) 12499 return SDValue(); 12500 12501 BitVector UndefElements; 12502 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(ConstVec); 12503 int32_t C = BV->getConstantFPSplatPow2ToLog2Int(&UndefElements, FloatBits + 1); 12504 if (C == -1 || C == 0 || C > FloatBits) 12505 return SDValue(); 12506 12507 MVT ResTy; 12508 unsigned NumLanes = Op.getValueType().getVectorNumElements(); 12509 switch (NumLanes) { 12510 default: 12511 return SDValue(); 12512 case 2: 12513 ResTy = FloatBits == 32 ? MVT::v2i32 : MVT::v2i64; 12514 break; 12515 case 4: 12516 ResTy = FloatBits == 32 ? MVT::v4i32 : MVT::v4i64; 12517 break; 12518 } 12519 12520 if (ResTy == MVT::v4i64 && DCI.isBeforeLegalizeOps()) 12521 return SDValue(); 12522 12523 SDLoc DL(N); 12524 SDValue ConvInput = Op.getOperand(0); 12525 bool IsSigned = Opc == ISD::SINT_TO_FP; 12526 if (IntBits < FloatBits) 12527 ConvInput = DAG.getNode(IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND, DL, 12528 ResTy, ConvInput); 12529 12530 unsigned IntrinsicOpcode = IsSigned ? Intrinsic::aarch64_neon_vcvtfxs2fp 12531 : Intrinsic::aarch64_neon_vcvtfxu2fp; 12532 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, Op.getValueType(), 12533 DAG.getConstant(IntrinsicOpcode, DL, MVT::i32), ConvInput, 12534 DAG.getConstant(C, DL, MVT::i32)); 12535 } 12536 12537 /// An EXTR instruction is made up of two shifts, ORed together. This helper 12538 /// searches for and classifies those shifts. 12539 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount, 12540 bool &FromHi) { 12541 if (N.getOpcode() == ISD::SHL) 12542 FromHi = false; 12543 else if (N.getOpcode() == ISD::SRL) 12544 FromHi = true; 12545 else 12546 return false; 12547 12548 if (!isa<ConstantSDNode>(N.getOperand(1))) 12549 return false; 12550 12551 ShiftAmount = N->getConstantOperandVal(1); 12552 Src = N->getOperand(0); 12553 return true; 12554 } 12555 12556 /// EXTR instruction extracts a contiguous chunk of bits from two existing 12557 /// registers viewed as a high/low pair. This function looks for the pattern: 12558 /// <tt>(or (shl VAL1, \#N), (srl VAL2, \#RegWidth-N))</tt> and replaces it 12559 /// with an EXTR. Can't quite be done in TableGen because the two immediates 12560 /// aren't independent. 12561 static SDValue tryCombineToEXTR(SDNode *N, 12562 TargetLowering::DAGCombinerInfo &DCI) { 12563 SelectionDAG &DAG = DCI.DAG; 12564 SDLoc DL(N); 12565 EVT VT = N->getValueType(0); 12566 12567 assert(N->getOpcode() == ISD::OR && "Unexpected root"); 12568 12569 if (VT != MVT::i32 && VT != MVT::i64) 12570 return SDValue(); 12571 12572 SDValue LHS; 12573 uint32_t ShiftLHS = 0; 12574 bool LHSFromHi = false; 12575 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi)) 12576 return SDValue(); 12577 12578 SDValue RHS; 12579 uint32_t ShiftRHS = 0; 12580 bool RHSFromHi = false; 12581 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi)) 12582 return SDValue(); 12583 12584 // If they're both trying to come from the high part of the register, they're 12585 // not really an EXTR. 12586 if (LHSFromHi == RHSFromHi) 12587 return SDValue(); 12588 12589 if (ShiftLHS + ShiftRHS != VT.getSizeInBits()) 12590 return SDValue(); 12591 12592 if (LHSFromHi) { 12593 std::swap(LHS, RHS); 12594 std::swap(ShiftLHS, ShiftRHS); 12595 } 12596 12597 return DAG.getNode(AArch64ISD::EXTR, DL, VT, LHS, RHS, 12598 DAG.getConstant(ShiftRHS, DL, MVT::i64)); 12599 } 12600 12601 static SDValue tryCombineToBSL(SDNode *N, 12602 TargetLowering::DAGCombinerInfo &DCI) { 12603 EVT VT = N->getValueType(0); 12604 SelectionDAG &DAG = DCI.DAG; 12605 SDLoc DL(N); 12606 12607 if (!VT.isVector()) 12608 return SDValue(); 12609 12610 // The combining code currently only works for NEON vectors. In particular, 12611 // it does not work for SVE when dealing with vectors wider than 128 bits. 12612 if (!VT.is64BitVector() && !VT.is128BitVector()) 12613 return SDValue(); 12614 12615 SDValue N0 = N->getOperand(0); 12616 if (N0.getOpcode() != ISD::AND) 12617 return SDValue(); 12618 12619 SDValue N1 = N->getOperand(1); 12620 if (N1.getOpcode() != ISD::AND) 12621 return SDValue(); 12622 12623 // InstCombine does (not (neg a)) => (add a -1). 12624 // Try: (or (and (neg a) b) (and (add a -1) c)) => (bsl (neg a) b c) 12625 // Loop over all combinations of AND operands. 12626 for (int i = 1; i >= 0; --i) { 12627 for (int j = 1; j >= 0; --j) { 12628 SDValue O0 = N0->getOperand(i); 12629 SDValue O1 = N1->getOperand(j); 12630 SDValue Sub, Add, SubSibling, AddSibling; 12631 12632 // Find a SUB and an ADD operand, one from each AND. 12633 if (O0.getOpcode() == ISD::SUB && O1.getOpcode() == ISD::ADD) { 12634 Sub = O0; 12635 Add = O1; 12636 SubSibling = N0->getOperand(1 - i); 12637 AddSibling = N1->getOperand(1 - j); 12638 } else if (O0.getOpcode() == ISD::ADD && O1.getOpcode() == ISD::SUB) { 12639 Add = O0; 12640 Sub = O1; 12641 AddSibling = N0->getOperand(1 - i); 12642 SubSibling = N1->getOperand(1 - j); 12643 } else 12644 continue; 12645 12646 if (!ISD::isBuildVectorAllZeros(Sub.getOperand(0).getNode())) 12647 continue; 12648 12649 // Constant ones is always righthand operand of the Add. 12650 if (!ISD::isBuildVectorAllOnes(Add.getOperand(1).getNode())) 12651 continue; 12652 12653 if (Sub.getOperand(1) != Add.getOperand(0)) 12654 continue; 12655 12656 return DAG.getNode(AArch64ISD::BSP, DL, VT, Sub, SubSibling, AddSibling); 12657 } 12658 } 12659 12660 // (or (and a b) (and (not a) c)) => (bsl a b c) 12661 // We only have to look for constant vectors here since the general, variable 12662 // case can be handled in TableGen. 12663 unsigned Bits = VT.getScalarSizeInBits(); 12664 uint64_t BitMask = Bits == 64 ? -1ULL : ((1ULL << Bits) - 1); 12665 for (int i = 1; i >= 0; --i) 12666 for (int j = 1; j >= 0; --j) { 12667 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(i)); 12668 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(j)); 12669 if (!BVN0 || !BVN1) 12670 continue; 12671 12672 bool FoundMatch = true; 12673 for (unsigned k = 0; k < VT.getVectorNumElements(); ++k) { 12674 ConstantSDNode *CN0 = dyn_cast<ConstantSDNode>(BVN0->getOperand(k)); 12675 ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(BVN1->getOperand(k)); 12676 if (!CN0 || !CN1 || 12677 CN0->getZExtValue() != (BitMask & ~CN1->getZExtValue())) { 12678 FoundMatch = false; 12679 break; 12680 } 12681 } 12682 12683 if (FoundMatch) 12684 return DAG.getNode(AArch64ISD::BSP, DL, VT, SDValue(BVN0, 0), 12685 N0->getOperand(1 - i), N1->getOperand(1 - j)); 12686 } 12687 12688 return SDValue(); 12689 } 12690 12691 static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI, 12692 const AArch64Subtarget *Subtarget) { 12693 // Attempt to form an EXTR from (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) 12694 SelectionDAG &DAG = DCI.DAG; 12695 EVT VT = N->getValueType(0); 12696 12697 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT)) 12698 return SDValue(); 12699 12700 if (SDValue Res = tryCombineToEXTR(N, DCI)) 12701 return Res; 12702 12703 if (SDValue Res = tryCombineToBSL(N, DCI)) 12704 return Res; 12705 12706 return SDValue(); 12707 } 12708 12709 static bool isConstantSplatVectorMaskForType(SDNode *N, EVT MemVT) { 12710 if (!MemVT.getVectorElementType().isSimple()) 12711 return false; 12712 12713 uint64_t MaskForTy = 0ull; 12714 switch (MemVT.getVectorElementType().getSimpleVT().SimpleTy) { 12715 case MVT::i8: 12716 MaskForTy = 0xffull; 12717 break; 12718 case MVT::i16: 12719 MaskForTy = 0xffffull; 12720 break; 12721 case MVT::i32: 12722 MaskForTy = 0xffffffffull; 12723 break; 12724 default: 12725 return false; 12726 break; 12727 } 12728 12729 if (N->getOpcode() == AArch64ISD::DUP || N->getOpcode() == ISD::SPLAT_VECTOR) 12730 if (auto *Op0 = dyn_cast<ConstantSDNode>(N->getOperand(0))) 12731 return Op0->getAPIntValue().getLimitedValue() == MaskForTy; 12732 12733 return false; 12734 } 12735 12736 static SDValue performSVEAndCombine(SDNode *N, 12737 TargetLowering::DAGCombinerInfo &DCI) { 12738 if (DCI.isBeforeLegalizeOps()) 12739 return SDValue(); 12740 12741 SelectionDAG &DAG = DCI.DAG; 12742 SDValue Src = N->getOperand(0); 12743 unsigned Opc = Src->getOpcode(); 12744 12745 // Zero/any extend of an unsigned unpack 12746 if (Opc == AArch64ISD::UUNPKHI || Opc == AArch64ISD::UUNPKLO) { 12747 SDValue UnpkOp = Src->getOperand(0); 12748 SDValue Dup = N->getOperand(1); 12749 12750 if (Dup.getOpcode() != AArch64ISD::DUP) 12751 return SDValue(); 12752 12753 SDLoc DL(N); 12754 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Dup->getOperand(0)); 12755 uint64_t ExtVal = C->getZExtValue(); 12756 12757 // If the mask is fully covered by the unpack, we don't need to push 12758 // a new AND onto the operand 12759 EVT EltTy = UnpkOp->getValueType(0).getVectorElementType(); 12760 if ((ExtVal == 0xFF && EltTy == MVT::i8) || 12761 (ExtVal == 0xFFFF && EltTy == MVT::i16) || 12762 (ExtVal == 0xFFFFFFFF && EltTy == MVT::i32)) 12763 return Src; 12764 12765 // Truncate to prevent a DUP with an over wide constant 12766 APInt Mask = C->getAPIntValue().trunc(EltTy.getSizeInBits()); 12767 12768 // Otherwise, make sure we propagate the AND to the operand 12769 // of the unpack 12770 Dup = DAG.getNode(AArch64ISD::DUP, DL, 12771 UnpkOp->getValueType(0), 12772 DAG.getConstant(Mask.zextOrTrunc(32), DL, MVT::i32)); 12773 12774 SDValue And = DAG.getNode(ISD::AND, DL, 12775 UnpkOp->getValueType(0), UnpkOp, Dup); 12776 12777 return DAG.getNode(Opc, DL, N->getValueType(0), And); 12778 } 12779 12780 if (!EnableCombineMGatherIntrinsics) 12781 return SDValue(); 12782 12783 SDValue Mask = N->getOperand(1); 12784 12785 if (!Src.hasOneUse()) 12786 return SDValue(); 12787 12788 EVT MemVT; 12789 12790 // SVE load instructions perform an implicit zero-extend, which makes them 12791 // perfect candidates for combining. 12792 switch (Opc) { 12793 case AArch64ISD::LD1_MERGE_ZERO: 12794 case AArch64ISD::LDNF1_MERGE_ZERO: 12795 case AArch64ISD::LDFF1_MERGE_ZERO: 12796 MemVT = cast<VTSDNode>(Src->getOperand(3))->getVT(); 12797 break; 12798 case AArch64ISD::GLD1_MERGE_ZERO: 12799 case AArch64ISD::GLD1_SCALED_MERGE_ZERO: 12800 case AArch64ISD::GLD1_SXTW_MERGE_ZERO: 12801 case AArch64ISD::GLD1_SXTW_SCALED_MERGE_ZERO: 12802 case AArch64ISD::GLD1_UXTW_MERGE_ZERO: 12803 case AArch64ISD::GLD1_UXTW_SCALED_MERGE_ZERO: 12804 case AArch64ISD::GLD1_IMM_MERGE_ZERO: 12805 case AArch64ISD::GLDFF1_MERGE_ZERO: 12806 case AArch64ISD::GLDFF1_SCALED_MERGE_ZERO: 12807 case AArch64ISD::GLDFF1_SXTW_MERGE_ZERO: 12808 case AArch64ISD::GLDFF1_SXTW_SCALED_MERGE_ZERO: 12809 case AArch64ISD::GLDFF1_UXTW_MERGE_ZERO: 12810 case AArch64ISD::GLDFF1_UXTW_SCALED_MERGE_ZERO: 12811 case AArch64ISD::GLDFF1_IMM_MERGE_ZERO: 12812 case AArch64ISD::GLDNT1_MERGE_ZERO: 12813 MemVT = cast<VTSDNode>(Src->getOperand(4))->getVT(); 12814 break; 12815 default: 12816 return SDValue(); 12817 } 12818 12819 if (isConstantSplatVectorMaskForType(Mask.getNode(), MemVT)) 12820 return Src; 12821 12822 return SDValue(); 12823 } 12824 12825 static SDValue performANDCombine(SDNode *N, 12826 TargetLowering::DAGCombinerInfo &DCI) { 12827 SelectionDAG &DAG = DCI.DAG; 12828 SDValue LHS = N->getOperand(0); 12829 EVT VT = N->getValueType(0); 12830 if (!VT.isVector() || !DAG.getTargetLoweringInfo().isTypeLegal(VT)) 12831 return SDValue(); 12832 12833 if (VT.isScalableVector()) 12834 return performSVEAndCombine(N, DCI); 12835 12836 // The combining code below works only for NEON vectors. In particular, it 12837 // does not work for SVE when dealing with vectors wider than 128 bits. 12838 if (!(VT.is64BitVector() || VT.is128BitVector())) 12839 return SDValue(); 12840 12841 BuildVectorSDNode *BVN = 12842 dyn_cast<BuildVectorSDNode>(N->getOperand(1).getNode()); 12843 if (!BVN) 12844 return SDValue(); 12845 12846 // AND does not accept an immediate, so check if we can use a BIC immediate 12847 // instruction instead. We do this here instead of using a (and x, (mvni imm)) 12848 // pattern in isel, because some immediates may be lowered to the preferred 12849 // (and x, (movi imm)) form, even though an mvni representation also exists. 12850 APInt DefBits(VT.getSizeInBits(), 0); 12851 APInt UndefBits(VT.getSizeInBits(), 0); 12852 if (resolveBuildVector(BVN, DefBits, UndefBits)) { 12853 SDValue NewOp; 12854 12855 DefBits = ~DefBits; 12856 if ((NewOp = tryAdvSIMDModImm32(AArch64ISD::BICi, SDValue(N, 0), DAG, 12857 DefBits, &LHS)) || 12858 (NewOp = tryAdvSIMDModImm16(AArch64ISD::BICi, SDValue(N, 0), DAG, 12859 DefBits, &LHS))) 12860 return NewOp; 12861 12862 UndefBits = ~UndefBits; 12863 if ((NewOp = tryAdvSIMDModImm32(AArch64ISD::BICi, SDValue(N, 0), DAG, 12864 UndefBits, &LHS)) || 12865 (NewOp = tryAdvSIMDModImm16(AArch64ISD::BICi, SDValue(N, 0), DAG, 12866 UndefBits, &LHS))) 12867 return NewOp; 12868 } 12869 12870 return SDValue(); 12871 } 12872 12873 static SDValue performSRLCombine(SDNode *N, 12874 TargetLowering::DAGCombinerInfo &DCI) { 12875 SelectionDAG &DAG = DCI.DAG; 12876 EVT VT = N->getValueType(0); 12877 if (VT != MVT::i32 && VT != MVT::i64) 12878 return SDValue(); 12879 12880 // Canonicalize (srl (bswap i32 x), 16) to (rotr (bswap i32 x), 16), if the 12881 // high 16-bits of x are zero. Similarly, canonicalize (srl (bswap i64 x), 32) 12882 // to (rotr (bswap i64 x), 32), if the high 32-bits of x are zero. 12883 SDValue N0 = N->getOperand(0); 12884 if (N0.getOpcode() == ISD::BSWAP) { 12885 SDLoc DL(N); 12886 SDValue N1 = N->getOperand(1); 12887 SDValue N00 = N0.getOperand(0); 12888 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) { 12889 uint64_t ShiftAmt = C->getZExtValue(); 12890 if (VT == MVT::i32 && ShiftAmt == 16 && 12891 DAG.MaskedValueIsZero(N00, APInt::getHighBitsSet(32, 16))) 12892 return DAG.getNode(ISD::ROTR, DL, VT, N0, N1); 12893 if (VT == MVT::i64 && ShiftAmt == 32 && 12894 DAG.MaskedValueIsZero(N00, APInt::getHighBitsSet(64, 32))) 12895 return DAG.getNode(ISD::ROTR, DL, VT, N0, N1); 12896 } 12897 } 12898 return SDValue(); 12899 } 12900 12901 // Attempt to form urhadd(OpA, OpB) from 12902 // truncate(vlshr(sub(zext(OpB), xor(zext(OpA), Ones(ElemSizeInBits))), 1)) 12903 // or uhadd(OpA, OpB) from truncate(vlshr(add(zext(OpA), zext(OpB)), 1)). 12904 // The original form of the first expression is 12905 // truncate(srl(add(zext(OpB), add(zext(OpA), 1)), 1)) and the 12906 // (OpA + OpB + 1) subexpression will have been changed to (OpB - (~OpA)). 12907 // Before this function is called the srl will have been lowered to 12908 // AArch64ISD::VLSHR. 12909 // This pass can also recognize signed variants of the patterns that use sign 12910 // extension instead of zero extension and form a srhadd(OpA, OpB) or a 12911 // shadd(OpA, OpB) from them. 12912 static SDValue 12913 performVectorTruncateCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI, 12914 SelectionDAG &DAG) { 12915 EVT VT = N->getValueType(0); 12916 12917 // Since we are looking for a right shift by a constant value of 1 and we are 12918 // operating on types at least 16 bits in length (sign/zero extended OpA and 12919 // OpB, which are at least 8 bits), it follows that the truncate will always 12920 // discard the shifted-in bit and therefore the right shift will be logical 12921 // regardless of the signedness of OpA and OpB. 12922 SDValue Shift = N->getOperand(0); 12923 if (Shift.getOpcode() != AArch64ISD::VLSHR) 12924 return SDValue(); 12925 12926 // Is the right shift using an immediate value of 1? 12927 uint64_t ShiftAmount = Shift.getConstantOperandVal(1); 12928 if (ShiftAmount != 1) 12929 return SDValue(); 12930 12931 SDValue ExtendOpA, ExtendOpB; 12932 SDValue ShiftOp0 = Shift.getOperand(0); 12933 unsigned ShiftOp0Opc = ShiftOp0.getOpcode(); 12934 if (ShiftOp0Opc == ISD::SUB) { 12935 12936 SDValue Xor = ShiftOp0.getOperand(1); 12937 if (Xor.getOpcode() != ISD::XOR) 12938 return SDValue(); 12939 12940 // Is the XOR using a constant amount of all ones in the right hand side? 12941 uint64_t C; 12942 if (!isAllConstantBuildVector(Xor.getOperand(1), C)) 12943 return SDValue(); 12944 12945 unsigned ElemSizeInBits = VT.getScalarSizeInBits(); 12946 APInt CAsAPInt(ElemSizeInBits, C); 12947 if (CAsAPInt != APInt::getAllOnesValue(ElemSizeInBits)) 12948 return SDValue(); 12949 12950 ExtendOpA = Xor.getOperand(0); 12951 ExtendOpB = ShiftOp0.getOperand(0); 12952 } else if (ShiftOp0Opc == ISD::ADD) { 12953 ExtendOpA = ShiftOp0.getOperand(0); 12954 ExtendOpB = ShiftOp0.getOperand(1); 12955 } else 12956 return SDValue(); 12957 12958 unsigned ExtendOpAOpc = ExtendOpA.getOpcode(); 12959 unsigned ExtendOpBOpc = ExtendOpB.getOpcode(); 12960 if (!(ExtendOpAOpc == ExtendOpBOpc && 12961 (ExtendOpAOpc == ISD::ZERO_EXTEND || ExtendOpAOpc == ISD::SIGN_EXTEND))) 12962 return SDValue(); 12963 12964 // Is the result of the right shift being truncated to the same value type as 12965 // the original operands, OpA and OpB? 12966 SDValue OpA = ExtendOpA.getOperand(0); 12967 SDValue OpB = ExtendOpB.getOperand(0); 12968 EVT OpAVT = OpA.getValueType(); 12969 assert(ExtendOpA.getValueType() == ExtendOpB.getValueType()); 12970 if (!(VT == OpAVT && OpAVT == OpB.getValueType())) 12971 return SDValue(); 12972 12973 SDLoc DL(N); 12974 bool IsSignExtend = ExtendOpAOpc == ISD::SIGN_EXTEND; 12975 bool IsRHADD = ShiftOp0Opc == ISD::SUB; 12976 unsigned HADDOpc = IsSignExtend 12977 ? (IsRHADD ? AArch64ISD::SRHADD : AArch64ISD::SHADD) 12978 : (IsRHADD ? AArch64ISD::URHADD : AArch64ISD::UHADD); 12979 SDValue ResultHADD = DAG.getNode(HADDOpc, DL, VT, OpA, OpB); 12980 12981 return ResultHADD; 12982 } 12983 12984 static bool hasPairwiseAdd(unsigned Opcode, EVT VT, bool FullFP16) { 12985 switch (Opcode) { 12986 case ISD::FADD: 12987 return (FullFP16 && VT == MVT::f16) || VT == MVT::f32 || VT == MVT::f64; 12988 case ISD::ADD: 12989 return VT == MVT::i64; 12990 default: 12991 return false; 12992 } 12993 } 12994 12995 static SDValue performExtractVectorEltCombine(SDNode *N, SelectionDAG &DAG) { 12996 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1); 12997 ConstantSDNode *ConstantN1 = dyn_cast<ConstantSDNode>(N1); 12998 12999 EVT VT = N->getValueType(0); 13000 const bool FullFP16 = 13001 static_cast<const AArch64Subtarget &>(DAG.getSubtarget()).hasFullFP16(); 13002 13003 // Rewrite for pairwise fadd pattern 13004 // (f32 (extract_vector_elt 13005 // (fadd (vXf32 Other) 13006 // (vector_shuffle (vXf32 Other) undef <1,X,...> )) 0)) 13007 // -> 13008 // (f32 (fadd (extract_vector_elt (vXf32 Other) 0) 13009 // (extract_vector_elt (vXf32 Other) 1)) 13010 if (ConstantN1 && ConstantN1->getZExtValue() == 0 && 13011 hasPairwiseAdd(N0->getOpcode(), VT, FullFP16)) { 13012 SDLoc DL(N0); 13013 SDValue N00 = N0->getOperand(0); 13014 SDValue N01 = N0->getOperand(1); 13015 13016 ShuffleVectorSDNode *Shuffle = dyn_cast<ShuffleVectorSDNode>(N01); 13017 SDValue Other = N00; 13018 13019 // And handle the commutative case. 13020 if (!Shuffle) { 13021 Shuffle = dyn_cast<ShuffleVectorSDNode>(N00); 13022 Other = N01; 13023 } 13024 13025 if (Shuffle && Shuffle->getMaskElt(0) == 1 && 13026 Other == Shuffle->getOperand(0)) { 13027 return DAG.getNode(N0->getOpcode(), DL, VT, 13028 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, Other, 13029 DAG.getConstant(0, DL, MVT::i64)), 13030 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, Other, 13031 DAG.getConstant(1, DL, MVT::i64))); 13032 } 13033 } 13034 13035 return SDValue(); 13036 } 13037 13038 static SDValue performConcatVectorsCombine(SDNode *N, 13039 TargetLowering::DAGCombinerInfo &DCI, 13040 SelectionDAG &DAG) { 13041 SDLoc dl(N); 13042 EVT VT = N->getValueType(0); 13043 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1); 13044 unsigned N0Opc = N0->getOpcode(), N1Opc = N1->getOpcode(); 13045 13046 // Optimize concat_vectors of truncated vectors, where the intermediate 13047 // type is illegal, to avoid said illegality, e.g., 13048 // (v4i16 (concat_vectors (v2i16 (truncate (v2i64))), 13049 // (v2i16 (truncate (v2i64))))) 13050 // -> 13051 // (v4i16 (truncate (vector_shuffle (v4i32 (bitcast (v2i64))), 13052 // (v4i32 (bitcast (v2i64))), 13053 // <0, 2, 4, 6>))) 13054 // This isn't really target-specific, but ISD::TRUNCATE legality isn't keyed 13055 // on both input and result type, so we might generate worse code. 13056 // On AArch64 we know it's fine for v2i64->v4i16 and v4i32->v8i8. 13057 if (N->getNumOperands() == 2 && N0Opc == ISD::TRUNCATE && 13058 N1Opc == ISD::TRUNCATE) { 13059 SDValue N00 = N0->getOperand(0); 13060 SDValue N10 = N1->getOperand(0); 13061 EVT N00VT = N00.getValueType(); 13062 13063 if (N00VT == N10.getValueType() && 13064 (N00VT == MVT::v2i64 || N00VT == MVT::v4i32) && 13065 N00VT.getScalarSizeInBits() == 4 * VT.getScalarSizeInBits()) { 13066 MVT MidVT = (N00VT == MVT::v2i64 ? MVT::v4i32 : MVT::v8i16); 13067 SmallVector<int, 8> Mask(MidVT.getVectorNumElements()); 13068 for (size_t i = 0; i < Mask.size(); ++i) 13069 Mask[i] = i * 2; 13070 return DAG.getNode(ISD::TRUNCATE, dl, VT, 13071 DAG.getVectorShuffle( 13072 MidVT, dl, 13073 DAG.getNode(ISD::BITCAST, dl, MidVT, N00), 13074 DAG.getNode(ISD::BITCAST, dl, MidVT, N10), Mask)); 13075 } 13076 } 13077 13078 // Wait 'til after everything is legalized to try this. That way we have 13079 // legal vector types and such. 13080 if (DCI.isBeforeLegalizeOps()) 13081 return SDValue(); 13082 13083 // Optimise concat_vectors of two [us]rhadds or [us]hadds that use extracted 13084 // subvectors from the same original vectors. Combine these into a single 13085 // [us]rhadd or [us]hadd that operates on the two original vectors. Example: 13086 // (v16i8 (concat_vectors (v8i8 (urhadd (extract_subvector (v16i8 OpA, <0>), 13087 // extract_subvector (v16i8 OpB, 13088 // <0>))), 13089 // (v8i8 (urhadd (extract_subvector (v16i8 OpA, <8>), 13090 // extract_subvector (v16i8 OpB, 13091 // <8>))))) 13092 // -> 13093 // (v16i8(urhadd(v16i8 OpA, v16i8 OpB))) 13094 if (N->getNumOperands() == 2 && N0Opc == N1Opc && 13095 (N0Opc == AArch64ISD::URHADD || N0Opc == AArch64ISD::SRHADD || 13096 N0Opc == AArch64ISD::UHADD || N0Opc == AArch64ISD::SHADD)) { 13097 SDValue N00 = N0->getOperand(0); 13098 SDValue N01 = N0->getOperand(1); 13099 SDValue N10 = N1->getOperand(0); 13100 SDValue N11 = N1->getOperand(1); 13101 13102 EVT N00VT = N00.getValueType(); 13103 EVT N10VT = N10.getValueType(); 13104 13105 if (N00->getOpcode() == ISD::EXTRACT_SUBVECTOR && 13106 N01->getOpcode() == ISD::EXTRACT_SUBVECTOR && 13107 N10->getOpcode() == ISD::EXTRACT_SUBVECTOR && 13108 N11->getOpcode() == ISD::EXTRACT_SUBVECTOR && N00VT == N10VT) { 13109 SDValue N00Source = N00->getOperand(0); 13110 SDValue N01Source = N01->getOperand(0); 13111 SDValue N10Source = N10->getOperand(0); 13112 SDValue N11Source = N11->getOperand(0); 13113 13114 if (N00Source == N10Source && N01Source == N11Source && 13115 N00Source.getValueType() == VT && N01Source.getValueType() == VT) { 13116 assert(N0.getValueType() == N1.getValueType()); 13117 13118 uint64_t N00Index = N00.getConstantOperandVal(1); 13119 uint64_t N01Index = N01.getConstantOperandVal(1); 13120 uint64_t N10Index = N10.getConstantOperandVal(1); 13121 uint64_t N11Index = N11.getConstantOperandVal(1); 13122 13123 if (N00Index == N01Index && N10Index == N11Index && N00Index == 0 && 13124 N10Index == N00VT.getVectorNumElements()) 13125 return DAG.getNode(N0Opc, dl, VT, N00Source, N01Source); 13126 } 13127 } 13128 } 13129 13130 // If we see a (concat_vectors (v1x64 A), (v1x64 A)) it's really a vector 13131 // splat. The indexed instructions are going to be expecting a DUPLANE64, so 13132 // canonicalise to that. 13133 if (N0 == N1 && VT.getVectorNumElements() == 2) { 13134 assert(VT.getScalarSizeInBits() == 64); 13135 return DAG.getNode(AArch64ISD::DUPLANE64, dl, VT, WidenVector(N0, DAG), 13136 DAG.getConstant(0, dl, MVT::i64)); 13137 } 13138 13139 // Canonicalise concat_vectors so that the right-hand vector has as few 13140 // bit-casts as possible before its real operation. The primary matching 13141 // destination for these operations will be the narrowing "2" instructions, 13142 // which depend on the operation being performed on this right-hand vector. 13143 // For example, 13144 // (concat_vectors LHS, (v1i64 (bitconvert (v4i16 RHS)))) 13145 // becomes 13146 // (bitconvert (concat_vectors (v4i16 (bitconvert LHS)), RHS)) 13147 13148 if (N1Opc != ISD::BITCAST) 13149 return SDValue(); 13150 SDValue RHS = N1->getOperand(0); 13151 MVT RHSTy = RHS.getValueType().getSimpleVT(); 13152 // If the RHS is not a vector, this is not the pattern we're looking for. 13153 if (!RHSTy.isVector()) 13154 return SDValue(); 13155 13156 LLVM_DEBUG( 13157 dbgs() << "aarch64-lower: concat_vectors bitcast simplification\n"); 13158 13159 MVT ConcatTy = MVT::getVectorVT(RHSTy.getVectorElementType(), 13160 RHSTy.getVectorNumElements() * 2); 13161 return DAG.getNode(ISD::BITCAST, dl, VT, 13162 DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatTy, 13163 DAG.getNode(ISD::BITCAST, dl, RHSTy, N0), 13164 RHS)); 13165 } 13166 13167 static SDValue tryCombineFixedPointConvert(SDNode *N, 13168 TargetLowering::DAGCombinerInfo &DCI, 13169 SelectionDAG &DAG) { 13170 // Wait until after everything is legalized to try this. That way we have 13171 // legal vector types and such. 13172 if (DCI.isBeforeLegalizeOps()) 13173 return SDValue(); 13174 // Transform a scalar conversion of a value from a lane extract into a 13175 // lane extract of a vector conversion. E.g., from foo1 to foo2: 13176 // double foo1(int64x2_t a) { return vcvtd_n_f64_s64(a[1], 9); } 13177 // double foo2(int64x2_t a) { return vcvtq_n_f64_s64(a, 9)[1]; } 13178 // 13179 // The second form interacts better with instruction selection and the 13180 // register allocator to avoid cross-class register copies that aren't 13181 // coalescable due to a lane reference. 13182 13183 // Check the operand and see if it originates from a lane extract. 13184 SDValue Op1 = N->getOperand(1); 13185 if (Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT) { 13186 // Yep, no additional predication needed. Perform the transform. 13187 SDValue IID = N->getOperand(0); 13188 SDValue Shift = N->getOperand(2); 13189 SDValue Vec = Op1.getOperand(0); 13190 SDValue Lane = Op1.getOperand(1); 13191 EVT ResTy = N->getValueType(0); 13192 EVT VecResTy; 13193 SDLoc DL(N); 13194 13195 // The vector width should be 128 bits by the time we get here, even 13196 // if it started as 64 bits (the extract_vector handling will have 13197 // done so). 13198 assert(Vec.getValueSizeInBits() == 128 && 13199 "unexpected vector size on extract_vector_elt!"); 13200 if (Vec.getValueType() == MVT::v4i32) 13201 VecResTy = MVT::v4f32; 13202 else if (Vec.getValueType() == MVT::v2i64) 13203 VecResTy = MVT::v2f64; 13204 else 13205 llvm_unreachable("unexpected vector type!"); 13206 13207 SDValue Convert = 13208 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VecResTy, IID, Vec, Shift); 13209 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResTy, Convert, Lane); 13210 } 13211 return SDValue(); 13212 } 13213 13214 // AArch64 high-vector "long" operations are formed by performing the non-high 13215 // version on an extract_subvector of each operand which gets the high half: 13216 // 13217 // (longop2 LHS, RHS) == (longop (extract_high LHS), (extract_high RHS)) 13218 // 13219 // However, there are cases which don't have an extract_high explicitly, but 13220 // have another operation that can be made compatible with one for free. For 13221 // example: 13222 // 13223 // (dupv64 scalar) --> (extract_high (dup128 scalar)) 13224 // 13225 // This routine does the actual conversion of such DUPs, once outer routines 13226 // have determined that everything else is in order. 13227 // It also supports immediate DUP-like nodes (MOVI/MVNi), which we can fold 13228 // similarly here. 13229 static SDValue tryExtendDUPToExtractHigh(SDValue N, SelectionDAG &DAG) { 13230 switch (N.getOpcode()) { 13231 case AArch64ISD::DUP: 13232 case AArch64ISD::DUPLANE8: 13233 case AArch64ISD::DUPLANE16: 13234 case AArch64ISD::DUPLANE32: 13235 case AArch64ISD::DUPLANE64: 13236 case AArch64ISD::MOVI: 13237 case AArch64ISD::MOVIshift: 13238 case AArch64ISD::MOVIedit: 13239 case AArch64ISD::MOVImsl: 13240 case AArch64ISD::MVNIshift: 13241 case AArch64ISD::MVNImsl: 13242 break; 13243 default: 13244 // FMOV could be supported, but isn't very useful, as it would only occur 13245 // if you passed a bitcast' floating point immediate to an eligible long 13246 // integer op (addl, smull, ...). 13247 return SDValue(); 13248 } 13249 13250 MVT NarrowTy = N.getSimpleValueType(); 13251 if (!NarrowTy.is64BitVector()) 13252 return SDValue(); 13253 13254 MVT ElementTy = NarrowTy.getVectorElementType(); 13255 unsigned NumElems = NarrowTy.getVectorNumElements(); 13256 MVT NewVT = MVT::getVectorVT(ElementTy, NumElems * 2); 13257 13258 SDLoc dl(N); 13259 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NarrowTy, 13260 DAG.getNode(N->getOpcode(), dl, NewVT, N->ops()), 13261 DAG.getConstant(NumElems, dl, MVT::i64)); 13262 } 13263 13264 static bool isEssentiallyExtractHighSubvector(SDValue N) { 13265 if (N.getOpcode() == ISD::BITCAST) 13266 N = N.getOperand(0); 13267 if (N.getOpcode() != ISD::EXTRACT_SUBVECTOR) 13268 return false; 13269 return cast<ConstantSDNode>(N.getOperand(1))->getAPIntValue() == 13270 N.getOperand(0).getValueType().getVectorNumElements() / 2; 13271 } 13272 13273 /// Helper structure to keep track of ISD::SET_CC operands. 13274 struct GenericSetCCInfo { 13275 const SDValue *Opnd0; 13276 const SDValue *Opnd1; 13277 ISD::CondCode CC; 13278 }; 13279 13280 /// Helper structure to keep track of a SET_CC lowered into AArch64 code. 13281 struct AArch64SetCCInfo { 13282 const SDValue *Cmp; 13283 AArch64CC::CondCode CC; 13284 }; 13285 13286 /// Helper structure to keep track of SetCC information. 13287 union SetCCInfo { 13288 GenericSetCCInfo Generic; 13289 AArch64SetCCInfo AArch64; 13290 }; 13291 13292 /// Helper structure to be able to read SetCC information. If set to 13293 /// true, IsAArch64 field, Info is a AArch64SetCCInfo, otherwise Info is a 13294 /// GenericSetCCInfo. 13295 struct SetCCInfoAndKind { 13296 SetCCInfo Info; 13297 bool IsAArch64; 13298 }; 13299 13300 /// Check whether or not \p Op is a SET_CC operation, either a generic or 13301 /// an 13302 /// AArch64 lowered one. 13303 /// \p SetCCInfo is filled accordingly. 13304 /// \post SetCCInfo is meanginfull only when this function returns true. 13305 /// \return True when Op is a kind of SET_CC operation. 13306 static bool isSetCC(SDValue Op, SetCCInfoAndKind &SetCCInfo) { 13307 // If this is a setcc, this is straight forward. 13308 if (Op.getOpcode() == ISD::SETCC) { 13309 SetCCInfo.Info.Generic.Opnd0 = &Op.getOperand(0); 13310 SetCCInfo.Info.Generic.Opnd1 = &Op.getOperand(1); 13311 SetCCInfo.Info.Generic.CC = cast<CondCodeSDNode>(Op.getOperand(2))->get(); 13312 SetCCInfo.IsAArch64 = false; 13313 return true; 13314 } 13315 // Otherwise, check if this is a matching csel instruction. 13316 // In other words: 13317 // - csel 1, 0, cc 13318 // - csel 0, 1, !cc 13319 if (Op.getOpcode() != AArch64ISD::CSEL) 13320 return false; 13321 // Set the information about the operands. 13322 // TODO: we want the operands of the Cmp not the csel 13323 SetCCInfo.Info.AArch64.Cmp = &Op.getOperand(3); 13324 SetCCInfo.IsAArch64 = true; 13325 SetCCInfo.Info.AArch64.CC = static_cast<AArch64CC::CondCode>( 13326 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue()); 13327 13328 // Check that the operands matches the constraints: 13329 // (1) Both operands must be constants. 13330 // (2) One must be 1 and the other must be 0. 13331 ConstantSDNode *TValue = dyn_cast<ConstantSDNode>(Op.getOperand(0)); 13332 ConstantSDNode *FValue = dyn_cast<ConstantSDNode>(Op.getOperand(1)); 13333 13334 // Check (1). 13335 if (!TValue || !FValue) 13336 return false; 13337 13338 // Check (2). 13339 if (!TValue->isOne()) { 13340 // Update the comparison when we are interested in !cc. 13341 std::swap(TValue, FValue); 13342 SetCCInfo.Info.AArch64.CC = 13343 AArch64CC::getInvertedCondCode(SetCCInfo.Info.AArch64.CC); 13344 } 13345 return TValue->isOne() && FValue->isNullValue(); 13346 } 13347 13348 // Returns true if Op is setcc or zext of setcc. 13349 static bool isSetCCOrZExtSetCC(const SDValue& Op, SetCCInfoAndKind &Info) { 13350 if (isSetCC(Op, Info)) 13351 return true; 13352 return ((Op.getOpcode() == ISD::ZERO_EXTEND) && 13353 isSetCC(Op->getOperand(0), Info)); 13354 } 13355 13356 // The folding we want to perform is: 13357 // (add x, [zext] (setcc cc ...) ) 13358 // --> 13359 // (csel x, (add x, 1), !cc ...) 13360 // 13361 // The latter will get matched to a CSINC instruction. 13362 static SDValue performSetccAddFolding(SDNode *Op, SelectionDAG &DAG) { 13363 assert(Op && Op->getOpcode() == ISD::ADD && "Unexpected operation!"); 13364 SDValue LHS = Op->getOperand(0); 13365 SDValue RHS = Op->getOperand(1); 13366 SetCCInfoAndKind InfoAndKind; 13367 13368 // If both operands are a SET_CC, then we don't want to perform this 13369 // folding and create another csel as this results in more instructions 13370 // (and higher register usage). 13371 if (isSetCCOrZExtSetCC(LHS, InfoAndKind) && 13372 isSetCCOrZExtSetCC(RHS, InfoAndKind)) 13373 return SDValue(); 13374 13375 // If neither operand is a SET_CC, give up. 13376 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind)) { 13377 std::swap(LHS, RHS); 13378 if (!isSetCCOrZExtSetCC(LHS, InfoAndKind)) 13379 return SDValue(); 13380 } 13381 13382 // FIXME: This could be generatized to work for FP comparisons. 13383 EVT CmpVT = InfoAndKind.IsAArch64 13384 ? InfoAndKind.Info.AArch64.Cmp->getOperand(0).getValueType() 13385 : InfoAndKind.Info.Generic.Opnd0->getValueType(); 13386 if (CmpVT != MVT::i32 && CmpVT != MVT::i64) 13387 return SDValue(); 13388 13389 SDValue CCVal; 13390 SDValue Cmp; 13391 SDLoc dl(Op); 13392 if (InfoAndKind.IsAArch64) { 13393 CCVal = DAG.getConstant( 13394 AArch64CC::getInvertedCondCode(InfoAndKind.Info.AArch64.CC), dl, 13395 MVT::i32); 13396 Cmp = *InfoAndKind.Info.AArch64.Cmp; 13397 } else 13398 Cmp = getAArch64Cmp( 13399 *InfoAndKind.Info.Generic.Opnd0, *InfoAndKind.Info.Generic.Opnd1, 13400 ISD::getSetCCInverse(InfoAndKind.Info.Generic.CC, CmpVT), CCVal, DAG, 13401 dl); 13402 13403 EVT VT = Op->getValueType(0); 13404 LHS = DAG.getNode(ISD::ADD, dl, VT, RHS, DAG.getConstant(1, dl, VT)); 13405 return DAG.getNode(AArch64ISD::CSEL, dl, VT, RHS, LHS, CCVal, Cmp); 13406 } 13407 13408 // ADD(UADDV a, UADDV b) --> UADDV(ADD a, b) 13409 static SDValue performUADDVCombine(SDNode *N, SelectionDAG &DAG) { 13410 EVT VT = N->getValueType(0); 13411 // Only scalar integer and vector types. 13412 if (N->getOpcode() != ISD::ADD || !VT.isScalarInteger()) 13413 return SDValue(); 13414 13415 SDValue LHS = N->getOperand(0); 13416 SDValue RHS = N->getOperand(1); 13417 if (LHS.getOpcode() != ISD::EXTRACT_VECTOR_ELT || 13418 RHS.getOpcode() != ISD::EXTRACT_VECTOR_ELT || LHS.getValueType() != VT) 13419 return SDValue(); 13420 13421 auto *LHSN1 = dyn_cast<ConstantSDNode>(LHS->getOperand(1)); 13422 auto *RHSN1 = dyn_cast<ConstantSDNode>(RHS->getOperand(1)); 13423 if (!LHSN1 || LHSN1 != RHSN1 || !RHSN1->isNullValue()) 13424 return SDValue(); 13425 13426 SDValue Op1 = LHS->getOperand(0); 13427 SDValue Op2 = RHS->getOperand(0); 13428 EVT OpVT1 = Op1.getValueType(); 13429 EVT OpVT2 = Op2.getValueType(); 13430 if (Op1.getOpcode() != AArch64ISD::UADDV || OpVT1 != OpVT2 || 13431 Op2.getOpcode() != AArch64ISD::UADDV || 13432 OpVT1.getVectorElementType() != VT) 13433 return SDValue(); 13434 13435 SDValue Val1 = Op1.getOperand(0); 13436 SDValue Val2 = Op2.getOperand(0); 13437 EVT ValVT = Val1->getValueType(0); 13438 SDLoc DL(N); 13439 SDValue AddVal = DAG.getNode(ISD::ADD, DL, ValVT, Val1, Val2); 13440 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, 13441 DAG.getNode(AArch64ISD::UADDV, DL, ValVT, AddVal), 13442 DAG.getConstant(0, DL, MVT::i64)); 13443 } 13444 13445 // ADD(UDOT(zero, x, y), A) --> UDOT(A, x, y) 13446 static SDValue performAddDotCombine(SDNode *N, SelectionDAG &DAG) { 13447 EVT VT = N->getValueType(0); 13448 if (N->getOpcode() != ISD::ADD) 13449 return SDValue(); 13450 13451 SDValue Dot = N->getOperand(0); 13452 SDValue A = N->getOperand(1); 13453 // Handle commutivity 13454 auto isZeroDot = [](SDValue Dot) { 13455 return (Dot.getOpcode() == AArch64ISD::UDOT || 13456 Dot.getOpcode() == AArch64ISD::SDOT) && 13457 isZerosVector(Dot.getOperand(0).getNode()); 13458 }; 13459 if (!isZeroDot(Dot)) 13460 std::swap(Dot, A); 13461 if (!isZeroDot(Dot)) 13462 return SDValue(); 13463 13464 return DAG.getNode(Dot.getOpcode(), SDLoc(N), VT, A, Dot.getOperand(1), 13465 Dot.getOperand(2)); 13466 } 13467 13468 // The basic add/sub long vector instructions have variants with "2" on the end 13469 // which act on the high-half of their inputs. They are normally matched by 13470 // patterns like: 13471 // 13472 // (add (zeroext (extract_high LHS)), 13473 // (zeroext (extract_high RHS))) 13474 // -> uaddl2 vD, vN, vM 13475 // 13476 // However, if one of the extracts is something like a duplicate, this 13477 // instruction can still be used profitably. This function puts the DAG into a 13478 // more appropriate form for those patterns to trigger. 13479 static SDValue performAddSubLongCombine(SDNode *N, 13480 TargetLowering::DAGCombinerInfo &DCI, 13481 SelectionDAG &DAG) { 13482 if (DCI.isBeforeLegalizeOps()) 13483 return SDValue(); 13484 13485 MVT VT = N->getSimpleValueType(0); 13486 if (!VT.is128BitVector()) { 13487 if (N->getOpcode() == ISD::ADD) 13488 return performSetccAddFolding(N, DAG); 13489 return SDValue(); 13490 } 13491 13492 // Make sure both branches are extended in the same way. 13493 SDValue LHS = N->getOperand(0); 13494 SDValue RHS = N->getOperand(1); 13495 if ((LHS.getOpcode() != ISD::ZERO_EXTEND && 13496 LHS.getOpcode() != ISD::SIGN_EXTEND) || 13497 LHS.getOpcode() != RHS.getOpcode()) 13498 return SDValue(); 13499 13500 unsigned ExtType = LHS.getOpcode(); 13501 13502 // It's not worth doing if at least one of the inputs isn't already an 13503 // extract, but we don't know which it'll be so we have to try both. 13504 if (isEssentiallyExtractHighSubvector(LHS.getOperand(0))) { 13505 RHS = tryExtendDUPToExtractHigh(RHS.getOperand(0), DAG); 13506 if (!RHS.getNode()) 13507 return SDValue(); 13508 13509 RHS = DAG.getNode(ExtType, SDLoc(N), VT, RHS); 13510 } else if (isEssentiallyExtractHighSubvector(RHS.getOperand(0))) { 13511 LHS = tryExtendDUPToExtractHigh(LHS.getOperand(0), DAG); 13512 if (!LHS.getNode()) 13513 return SDValue(); 13514 13515 LHS = DAG.getNode(ExtType, SDLoc(N), VT, LHS); 13516 } 13517 13518 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, LHS, RHS); 13519 } 13520 13521 static SDValue performAddSubCombine(SDNode *N, 13522 TargetLowering::DAGCombinerInfo &DCI, 13523 SelectionDAG &DAG) { 13524 // Try to change sum of two reductions. 13525 if (SDValue Val = performUADDVCombine(N, DAG)) 13526 return Val; 13527 if (SDValue Val = performAddDotCombine(N, DAG)) 13528 return Val; 13529 13530 return performAddSubLongCombine(N, DCI, DAG); 13531 } 13532 13533 // Massage DAGs which we can use the high-half "long" operations on into 13534 // something isel will recognize better. E.g. 13535 // 13536 // (aarch64_neon_umull (extract_high vec) (dupv64 scalar)) --> 13537 // (aarch64_neon_umull (extract_high (v2i64 vec))) 13538 // (extract_high (v2i64 (dup128 scalar))))) 13539 // 13540 static SDValue tryCombineLongOpWithDup(unsigned IID, SDNode *N, 13541 TargetLowering::DAGCombinerInfo &DCI, 13542 SelectionDAG &DAG) { 13543 if (DCI.isBeforeLegalizeOps()) 13544 return SDValue(); 13545 13546 SDValue LHS = N->getOperand((IID == Intrinsic::not_intrinsic) ? 0 : 1); 13547 SDValue RHS = N->getOperand((IID == Intrinsic::not_intrinsic) ? 1 : 2); 13548 assert(LHS.getValueType().is64BitVector() && 13549 RHS.getValueType().is64BitVector() && 13550 "unexpected shape for long operation"); 13551 13552 // Either node could be a DUP, but it's not worth doing both of them (you'd 13553 // just as well use the non-high version) so look for a corresponding extract 13554 // operation on the other "wing". 13555 if (isEssentiallyExtractHighSubvector(LHS)) { 13556 RHS = tryExtendDUPToExtractHigh(RHS, DAG); 13557 if (!RHS.getNode()) 13558 return SDValue(); 13559 } else if (isEssentiallyExtractHighSubvector(RHS)) { 13560 LHS = tryExtendDUPToExtractHigh(LHS, DAG); 13561 if (!LHS.getNode()) 13562 return SDValue(); 13563 } 13564 13565 if (IID == Intrinsic::not_intrinsic) 13566 return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0), LHS, RHS); 13567 13568 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0), 13569 N->getOperand(0), LHS, RHS); 13570 } 13571 13572 static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) { 13573 MVT ElemTy = N->getSimpleValueType(0).getScalarType(); 13574 unsigned ElemBits = ElemTy.getSizeInBits(); 13575 13576 int64_t ShiftAmount; 13577 if (BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(2))) { 13578 APInt SplatValue, SplatUndef; 13579 unsigned SplatBitSize; 13580 bool HasAnyUndefs; 13581 if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, 13582 HasAnyUndefs, ElemBits) || 13583 SplatBitSize != ElemBits) 13584 return SDValue(); 13585 13586 ShiftAmount = SplatValue.getSExtValue(); 13587 } else if (ConstantSDNode *CVN = dyn_cast<ConstantSDNode>(N->getOperand(2))) { 13588 ShiftAmount = CVN->getSExtValue(); 13589 } else 13590 return SDValue(); 13591 13592 unsigned Opcode; 13593 bool IsRightShift; 13594 switch (IID) { 13595 default: 13596 llvm_unreachable("Unknown shift intrinsic"); 13597 case Intrinsic::aarch64_neon_sqshl: 13598 Opcode = AArch64ISD::SQSHL_I; 13599 IsRightShift = false; 13600 break; 13601 case Intrinsic::aarch64_neon_uqshl: 13602 Opcode = AArch64ISD::UQSHL_I; 13603 IsRightShift = false; 13604 break; 13605 case Intrinsic::aarch64_neon_srshl: 13606 Opcode = AArch64ISD::SRSHR_I; 13607 IsRightShift = true; 13608 break; 13609 case Intrinsic::aarch64_neon_urshl: 13610 Opcode = AArch64ISD::URSHR_I; 13611 IsRightShift = true; 13612 break; 13613 case Intrinsic::aarch64_neon_sqshlu: 13614 Opcode = AArch64ISD::SQSHLU_I; 13615 IsRightShift = false; 13616 break; 13617 case Intrinsic::aarch64_neon_sshl: 13618 case Intrinsic::aarch64_neon_ushl: 13619 // For positive shift amounts we can use SHL, as ushl/sshl perform a regular 13620 // left shift for positive shift amounts. Below, we only replace the current 13621 // node with VSHL, if this condition is met. 13622 Opcode = AArch64ISD::VSHL; 13623 IsRightShift = false; 13624 break; 13625 } 13626 13627 if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits) { 13628 SDLoc dl(N); 13629 return DAG.getNode(Opcode, dl, N->getValueType(0), N->getOperand(1), 13630 DAG.getConstant(-ShiftAmount, dl, MVT::i32)); 13631 } else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount < ElemBits) { 13632 SDLoc dl(N); 13633 return DAG.getNode(Opcode, dl, N->getValueType(0), N->getOperand(1), 13634 DAG.getConstant(ShiftAmount, dl, MVT::i32)); 13635 } 13636 13637 return SDValue(); 13638 } 13639 13640 // The CRC32[BH] instructions ignore the high bits of their data operand. Since 13641 // the intrinsics must be legal and take an i32, this means there's almost 13642 // certainly going to be a zext in the DAG which we can eliminate. 13643 static SDValue tryCombineCRC32(unsigned Mask, SDNode *N, SelectionDAG &DAG) { 13644 SDValue AndN = N->getOperand(2); 13645 if (AndN.getOpcode() != ISD::AND) 13646 return SDValue(); 13647 13648 ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(AndN.getOperand(1)); 13649 if (!CMask || CMask->getZExtValue() != Mask) 13650 return SDValue(); 13651 13652 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), MVT::i32, 13653 N->getOperand(0), N->getOperand(1), AndN.getOperand(0)); 13654 } 13655 13656 static SDValue combineAcrossLanesIntrinsic(unsigned Opc, SDNode *N, 13657 SelectionDAG &DAG) { 13658 SDLoc dl(N); 13659 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), 13660 DAG.getNode(Opc, dl, 13661 N->getOperand(1).getSimpleValueType(), 13662 N->getOperand(1)), 13663 DAG.getConstant(0, dl, MVT::i64)); 13664 } 13665 13666 static SDValue LowerSVEIntrinsicIndex(SDNode *N, SelectionDAG &DAG) { 13667 SDLoc DL(N); 13668 SDValue Op1 = N->getOperand(1); 13669 SDValue Op2 = N->getOperand(2); 13670 EVT ScalarTy = Op2.getValueType(); 13671 if ((ScalarTy == MVT::i8) || (ScalarTy == MVT::i16)) 13672 ScalarTy = MVT::i32; 13673 13674 // Lower index_vector(base, step) to mul(step step_vector(1)) + splat(base). 13675 SDValue One = DAG.getConstant(1, DL, ScalarTy); 13676 SDValue StepVector = 13677 DAG.getNode(ISD::STEP_VECTOR, DL, N->getValueType(0), One); 13678 SDValue Step = DAG.getNode(ISD::SPLAT_VECTOR, DL, N->getValueType(0), Op2); 13679 SDValue Mul = DAG.getNode(ISD::MUL, DL, N->getValueType(0), StepVector, Step); 13680 SDValue Base = DAG.getNode(ISD::SPLAT_VECTOR, DL, N->getValueType(0), Op1); 13681 return DAG.getNode(ISD::ADD, DL, N->getValueType(0), Mul, Base); 13682 } 13683 13684 static SDValue LowerSVEIntrinsicDUP(SDNode *N, SelectionDAG &DAG) { 13685 SDLoc dl(N); 13686 SDValue Scalar = N->getOperand(3); 13687 EVT ScalarTy = Scalar.getValueType(); 13688 13689 if ((ScalarTy == MVT::i8) || (ScalarTy == MVT::i16)) 13690 Scalar = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Scalar); 13691 13692 SDValue Passthru = N->getOperand(1); 13693 SDValue Pred = N->getOperand(2); 13694 return DAG.getNode(AArch64ISD::DUP_MERGE_PASSTHRU, dl, N->getValueType(0), 13695 Pred, Scalar, Passthru); 13696 } 13697 13698 static SDValue LowerSVEIntrinsicEXT(SDNode *N, SelectionDAG &DAG) { 13699 SDLoc dl(N); 13700 LLVMContext &Ctx = *DAG.getContext(); 13701 EVT VT = N->getValueType(0); 13702 13703 assert(VT.isScalableVector() && "Expected a scalable vector."); 13704 13705 // Current lowering only supports the SVE-ACLE types. 13706 if (VT.getSizeInBits().getKnownMinSize() != AArch64::SVEBitsPerBlock) 13707 return SDValue(); 13708 13709 unsigned ElemSize = VT.getVectorElementType().getSizeInBits() / 8; 13710 unsigned ByteSize = VT.getSizeInBits().getKnownMinSize() / 8; 13711 EVT ByteVT = 13712 EVT::getVectorVT(Ctx, MVT::i8, ElementCount::getScalable(ByteSize)); 13713 13714 // Convert everything to the domain of EXT (i.e bytes). 13715 SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, ByteVT, N->getOperand(1)); 13716 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, ByteVT, N->getOperand(2)); 13717 SDValue Op2 = DAG.getNode(ISD::MUL, dl, MVT::i32, N->getOperand(3), 13718 DAG.getConstant(ElemSize, dl, MVT::i32)); 13719 13720 SDValue EXT = DAG.getNode(AArch64ISD::EXT, dl, ByteVT, Op0, Op1, Op2); 13721 return DAG.getNode(ISD::BITCAST, dl, VT, EXT); 13722 } 13723 13724 static SDValue tryConvertSVEWideCompare(SDNode *N, ISD::CondCode CC, 13725 TargetLowering::DAGCombinerInfo &DCI, 13726 SelectionDAG &DAG) { 13727 if (DCI.isBeforeLegalize()) 13728 return SDValue(); 13729 13730 SDValue Comparator = N->getOperand(3); 13731 if (Comparator.getOpcode() == AArch64ISD::DUP || 13732 Comparator.getOpcode() == ISD::SPLAT_VECTOR) { 13733 unsigned IID = getIntrinsicID(N); 13734 EVT VT = N->getValueType(0); 13735 EVT CmpVT = N->getOperand(2).getValueType(); 13736 SDValue Pred = N->getOperand(1); 13737 SDValue Imm; 13738 SDLoc DL(N); 13739 13740 switch (IID) { 13741 default: 13742 llvm_unreachable("Called with wrong intrinsic!"); 13743 break; 13744 13745 // Signed comparisons 13746 case Intrinsic::aarch64_sve_cmpeq_wide: 13747 case Intrinsic::aarch64_sve_cmpne_wide: 13748 case Intrinsic::aarch64_sve_cmpge_wide: 13749 case Intrinsic::aarch64_sve_cmpgt_wide: 13750 case Intrinsic::aarch64_sve_cmplt_wide: 13751 case Intrinsic::aarch64_sve_cmple_wide: { 13752 if (auto *CN = dyn_cast<ConstantSDNode>(Comparator.getOperand(0))) { 13753 int64_t ImmVal = CN->getSExtValue(); 13754 if (ImmVal >= -16 && ImmVal <= 15) 13755 Imm = DAG.getConstant(ImmVal, DL, MVT::i32); 13756 else 13757 return SDValue(); 13758 } 13759 break; 13760 } 13761 // Unsigned comparisons 13762 case Intrinsic::aarch64_sve_cmphs_wide: 13763 case Intrinsic::aarch64_sve_cmphi_wide: 13764 case Intrinsic::aarch64_sve_cmplo_wide: 13765 case Intrinsic::aarch64_sve_cmpls_wide: { 13766 if (auto *CN = dyn_cast<ConstantSDNode>(Comparator.getOperand(0))) { 13767 uint64_t ImmVal = CN->getZExtValue(); 13768 if (ImmVal <= 127) 13769 Imm = DAG.getConstant(ImmVal, DL, MVT::i32); 13770 else 13771 return SDValue(); 13772 } 13773 break; 13774 } 13775 } 13776 13777 if (!Imm) 13778 return SDValue(); 13779 13780 SDValue Splat = DAG.getNode(ISD::SPLAT_VECTOR, DL, CmpVT, Imm); 13781 return DAG.getNode(AArch64ISD::SETCC_MERGE_ZERO, DL, VT, Pred, 13782 N->getOperand(2), Splat, DAG.getCondCode(CC)); 13783 } 13784 13785 return SDValue(); 13786 } 13787 13788 static SDValue getPTest(SelectionDAG &DAG, EVT VT, SDValue Pg, SDValue Op, 13789 AArch64CC::CondCode Cond) { 13790 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 13791 13792 SDLoc DL(Op); 13793 assert(Op.getValueType().isScalableVector() && 13794 TLI.isTypeLegal(Op.getValueType()) && 13795 "Expected legal scalable vector type!"); 13796 13797 // Ensure target specific opcodes are using legal type. 13798 EVT OutVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT); 13799 SDValue TVal = DAG.getConstant(1, DL, OutVT); 13800 SDValue FVal = DAG.getConstant(0, DL, OutVT); 13801 13802 // Set condition code (CC) flags. 13803 SDValue Test = DAG.getNode(AArch64ISD::PTEST, DL, MVT::Other, Pg, Op); 13804 13805 // Convert CC to integer based on requested condition. 13806 // NOTE: Cond is inverted to promote CSEL's removal when it feeds a compare. 13807 SDValue CC = DAG.getConstant(getInvertedCondCode(Cond), DL, MVT::i32); 13808 SDValue Res = DAG.getNode(AArch64ISD::CSEL, DL, OutVT, FVal, TVal, CC, Test); 13809 return DAG.getZExtOrTrunc(Res, DL, VT); 13810 } 13811 13812 static SDValue combineSVEReductionInt(SDNode *N, unsigned Opc, 13813 SelectionDAG &DAG) { 13814 SDLoc DL(N); 13815 13816 SDValue Pred = N->getOperand(1); 13817 SDValue VecToReduce = N->getOperand(2); 13818 13819 // NOTE: The integer reduction's result type is not always linked to the 13820 // operand's element type so we construct it from the intrinsic's result type. 13821 EVT ReduceVT = getPackedSVEVectorVT(N->getValueType(0)); 13822 SDValue Reduce = DAG.getNode(Opc, DL, ReduceVT, Pred, VecToReduce); 13823 13824 // SVE reductions set the whole vector register with the first element 13825 // containing the reduction result, which we'll now extract. 13826 SDValue Zero = DAG.getConstant(0, DL, MVT::i64); 13827 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, N->getValueType(0), Reduce, 13828 Zero); 13829 } 13830 13831 static SDValue combineSVEReductionFP(SDNode *N, unsigned Opc, 13832 SelectionDAG &DAG) { 13833 SDLoc DL(N); 13834 13835 SDValue Pred = N->getOperand(1); 13836 SDValue VecToReduce = N->getOperand(2); 13837 13838 EVT ReduceVT = VecToReduce.getValueType(); 13839 SDValue Reduce = DAG.getNode(Opc, DL, ReduceVT, Pred, VecToReduce); 13840 13841 // SVE reductions set the whole vector register with the first element 13842 // containing the reduction result, which we'll now extract. 13843 SDValue Zero = DAG.getConstant(0, DL, MVT::i64); 13844 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, N->getValueType(0), Reduce, 13845 Zero); 13846 } 13847 13848 static SDValue combineSVEReductionOrderedFP(SDNode *N, unsigned Opc, 13849 SelectionDAG &DAG) { 13850 SDLoc DL(N); 13851 13852 SDValue Pred = N->getOperand(1); 13853 SDValue InitVal = N->getOperand(2); 13854 SDValue VecToReduce = N->getOperand(3); 13855 EVT ReduceVT = VecToReduce.getValueType(); 13856 13857 // Ordered reductions use the first lane of the result vector as the 13858 // reduction's initial value. 13859 SDValue Zero = DAG.getConstant(0, DL, MVT::i64); 13860 InitVal = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ReduceVT, 13861 DAG.getUNDEF(ReduceVT), InitVal, Zero); 13862 13863 SDValue Reduce = DAG.getNode(Opc, DL, ReduceVT, Pred, InitVal, VecToReduce); 13864 13865 // SVE reductions set the whole vector register with the first element 13866 // containing the reduction result, which we'll now extract. 13867 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, N->getValueType(0), Reduce, 13868 Zero); 13869 } 13870 13871 static bool isAllActivePredicate(SDValue N) { 13872 unsigned NumElts = N.getValueType().getVectorMinNumElements(); 13873 13874 // Look through cast. 13875 while (N.getOpcode() == AArch64ISD::REINTERPRET_CAST) { 13876 N = N.getOperand(0); 13877 // When reinterpreting from a type with fewer elements the "new" elements 13878 // are not active, so bail if they're likely to be used. 13879 if (N.getValueType().getVectorMinNumElements() < NumElts) 13880 return false; 13881 } 13882 13883 // "ptrue p.<ty>, all" can be considered all active when <ty> is the same size 13884 // or smaller than the implicit element type represented by N. 13885 // NOTE: A larger element count implies a smaller element type. 13886 if (N.getOpcode() == AArch64ISD::PTRUE && 13887 N.getConstantOperandVal(0) == AArch64SVEPredPattern::all) 13888 return N.getValueType().getVectorMinNumElements() >= NumElts; 13889 13890 return false; 13891 } 13892 13893 // If a merged operation has no inactive lanes we can relax it to a predicated 13894 // or unpredicated operation, which potentially allows better isel (perhaps 13895 // using immediate forms) or relaxing register reuse requirements. 13896 static SDValue convertMergedOpToPredOp(SDNode *N, unsigned Opc, 13897 SelectionDAG &DAG, 13898 bool UnpredOp = false) { 13899 assert(N->getOpcode() == ISD::INTRINSIC_WO_CHAIN && "Expected intrinsic!"); 13900 assert(N->getNumOperands() == 4 && "Expected 3 operand intrinsic!"); 13901 SDValue Pg = N->getOperand(1); 13902 13903 // ISD way to specify an all active predicate. 13904 if (isAllActivePredicate(Pg)) { 13905 if (UnpredOp) 13906 return DAG.getNode(Opc, SDLoc(N), N->getValueType(0), N->getOperand(2), 13907 N->getOperand(3)); 13908 else 13909 return DAG.getNode(Opc, SDLoc(N), N->getValueType(0), Pg, 13910 N->getOperand(2), N->getOperand(3)); 13911 } 13912 13913 // FUTURE: SplatVector(true) 13914 return SDValue(); 13915 } 13916 13917 static SDValue performIntrinsicCombine(SDNode *N, 13918 TargetLowering::DAGCombinerInfo &DCI, 13919 const AArch64Subtarget *Subtarget) { 13920 SelectionDAG &DAG = DCI.DAG; 13921 unsigned IID = getIntrinsicID(N); 13922 switch (IID) { 13923 default: 13924 break; 13925 case Intrinsic::aarch64_neon_vcvtfxs2fp: 13926 case Intrinsic::aarch64_neon_vcvtfxu2fp: 13927 return tryCombineFixedPointConvert(N, DCI, DAG); 13928 case Intrinsic::aarch64_neon_saddv: 13929 return combineAcrossLanesIntrinsic(AArch64ISD::SADDV, N, DAG); 13930 case Intrinsic::aarch64_neon_uaddv: 13931 return combineAcrossLanesIntrinsic(AArch64ISD::UADDV, N, DAG); 13932 case Intrinsic::aarch64_neon_sminv: 13933 return combineAcrossLanesIntrinsic(AArch64ISD::SMINV, N, DAG); 13934 case Intrinsic::aarch64_neon_uminv: 13935 return combineAcrossLanesIntrinsic(AArch64ISD::UMINV, N, DAG); 13936 case Intrinsic::aarch64_neon_smaxv: 13937 return combineAcrossLanesIntrinsic(AArch64ISD::SMAXV, N, DAG); 13938 case Intrinsic::aarch64_neon_umaxv: 13939 return combineAcrossLanesIntrinsic(AArch64ISD::UMAXV, N, DAG); 13940 case Intrinsic::aarch64_neon_fmax: 13941 return DAG.getNode(ISD::FMAXIMUM, SDLoc(N), N->getValueType(0), 13942 N->getOperand(1), N->getOperand(2)); 13943 case Intrinsic::aarch64_neon_fmin: 13944 return DAG.getNode(ISD::FMINIMUM, SDLoc(N), N->getValueType(0), 13945 N->getOperand(1), N->getOperand(2)); 13946 case Intrinsic::aarch64_neon_fmaxnm: 13947 return DAG.getNode(ISD::FMAXNUM, SDLoc(N), N->getValueType(0), 13948 N->getOperand(1), N->getOperand(2)); 13949 case Intrinsic::aarch64_neon_fminnm: 13950 return DAG.getNode(ISD::FMINNUM, SDLoc(N), N->getValueType(0), 13951 N->getOperand(1), N->getOperand(2)); 13952 case Intrinsic::aarch64_neon_smull: 13953 case Intrinsic::aarch64_neon_umull: 13954 case Intrinsic::aarch64_neon_pmull: 13955 case Intrinsic::aarch64_neon_sqdmull: 13956 return tryCombineLongOpWithDup(IID, N, DCI, DAG); 13957 case Intrinsic::aarch64_neon_sqshl: 13958 case Intrinsic::aarch64_neon_uqshl: 13959 case Intrinsic::aarch64_neon_sqshlu: 13960 case Intrinsic::aarch64_neon_srshl: 13961 case Intrinsic::aarch64_neon_urshl: 13962 case Intrinsic::aarch64_neon_sshl: 13963 case Intrinsic::aarch64_neon_ushl: 13964 return tryCombineShiftImm(IID, N, DAG); 13965 case Intrinsic::aarch64_crc32b: 13966 case Intrinsic::aarch64_crc32cb: 13967 return tryCombineCRC32(0xff, N, DAG); 13968 case Intrinsic::aarch64_crc32h: 13969 case Intrinsic::aarch64_crc32ch: 13970 return tryCombineCRC32(0xffff, N, DAG); 13971 case Intrinsic::aarch64_sve_saddv: 13972 // There is no i64 version of SADDV because the sign is irrelevant. 13973 if (N->getOperand(2)->getValueType(0).getVectorElementType() == MVT::i64) 13974 return combineSVEReductionInt(N, AArch64ISD::UADDV_PRED, DAG); 13975 else 13976 return combineSVEReductionInt(N, AArch64ISD::SADDV_PRED, DAG); 13977 case Intrinsic::aarch64_sve_uaddv: 13978 return combineSVEReductionInt(N, AArch64ISD::UADDV_PRED, DAG); 13979 case Intrinsic::aarch64_sve_smaxv: 13980 return combineSVEReductionInt(N, AArch64ISD::SMAXV_PRED, DAG); 13981 case Intrinsic::aarch64_sve_umaxv: 13982 return combineSVEReductionInt(N, AArch64ISD::UMAXV_PRED, DAG); 13983 case Intrinsic::aarch64_sve_sminv: 13984 return combineSVEReductionInt(N, AArch64ISD::SMINV_PRED, DAG); 13985 case Intrinsic::aarch64_sve_uminv: 13986 return combineSVEReductionInt(N, AArch64ISD::UMINV_PRED, DAG); 13987 case Intrinsic::aarch64_sve_orv: 13988 return combineSVEReductionInt(N, AArch64ISD::ORV_PRED, DAG); 13989 case Intrinsic::aarch64_sve_eorv: 13990 return combineSVEReductionInt(N, AArch64ISD::EORV_PRED, DAG); 13991 case Intrinsic::aarch64_sve_andv: 13992 return combineSVEReductionInt(N, AArch64ISD::ANDV_PRED, DAG); 13993 case Intrinsic::aarch64_sve_index: 13994 return LowerSVEIntrinsicIndex(N, DAG); 13995 case Intrinsic::aarch64_sve_dup: 13996 return LowerSVEIntrinsicDUP(N, DAG); 13997 case Intrinsic::aarch64_sve_dup_x: 13998 return DAG.getNode(ISD::SPLAT_VECTOR, SDLoc(N), N->getValueType(0), 13999 N->getOperand(1)); 14000 case Intrinsic::aarch64_sve_ext: 14001 return LowerSVEIntrinsicEXT(N, DAG); 14002 case Intrinsic::aarch64_sve_mul: 14003 return convertMergedOpToPredOp(N, AArch64ISD::MUL_PRED, DAG); 14004 case Intrinsic::aarch64_sve_smulh: 14005 return convertMergedOpToPredOp(N, AArch64ISD::MULHS_PRED, DAG); 14006 case Intrinsic::aarch64_sve_umulh: 14007 return convertMergedOpToPredOp(N, AArch64ISD::MULHU_PRED, DAG); 14008 case Intrinsic::aarch64_sve_smin: 14009 return convertMergedOpToPredOp(N, AArch64ISD::SMIN_PRED, DAG); 14010 case Intrinsic::aarch64_sve_umin: 14011 return convertMergedOpToPredOp(N, AArch64ISD::UMIN_PRED, DAG); 14012 case Intrinsic::aarch64_sve_smax: 14013 return convertMergedOpToPredOp(N, AArch64ISD::SMAX_PRED, DAG); 14014 case Intrinsic::aarch64_sve_umax: 14015 return convertMergedOpToPredOp(N, AArch64ISD::UMAX_PRED, DAG); 14016 case Intrinsic::aarch64_sve_lsl: 14017 return convertMergedOpToPredOp(N, AArch64ISD::SHL_PRED, DAG); 14018 case Intrinsic::aarch64_sve_lsr: 14019 return convertMergedOpToPredOp(N, AArch64ISD::SRL_PRED, DAG); 14020 case Intrinsic::aarch64_sve_asr: 14021 return convertMergedOpToPredOp(N, AArch64ISD::SRA_PRED, DAG); 14022 case Intrinsic::aarch64_sve_fadd: 14023 return convertMergedOpToPredOp(N, AArch64ISD::FADD_PRED, DAG); 14024 case Intrinsic::aarch64_sve_fsub: 14025 return convertMergedOpToPredOp(N, AArch64ISD::FSUB_PRED, DAG); 14026 case Intrinsic::aarch64_sve_fmul: 14027 return convertMergedOpToPredOp(N, AArch64ISD::FMUL_PRED, DAG); 14028 case Intrinsic::aarch64_sve_add: 14029 return convertMergedOpToPredOp(N, ISD::ADD, DAG, true); 14030 case Intrinsic::aarch64_sve_sub: 14031 return convertMergedOpToPredOp(N, ISD::SUB, DAG, true); 14032 case Intrinsic::aarch64_sve_and: 14033 return convertMergedOpToPredOp(N, ISD::AND, DAG, true); 14034 case Intrinsic::aarch64_sve_bic: 14035 return convertMergedOpToPredOp(N, AArch64ISD::BIC, DAG, true); 14036 case Intrinsic::aarch64_sve_eor: 14037 return convertMergedOpToPredOp(N, ISD::XOR, DAG, true); 14038 case Intrinsic::aarch64_sve_orr: 14039 return convertMergedOpToPredOp(N, ISD::OR, DAG, true); 14040 case Intrinsic::aarch64_sve_sqadd: 14041 return convertMergedOpToPredOp(N, ISD::SADDSAT, DAG, true); 14042 case Intrinsic::aarch64_sve_sqsub: 14043 return convertMergedOpToPredOp(N, ISD::SSUBSAT, DAG, true); 14044 case Intrinsic::aarch64_sve_uqadd: 14045 return convertMergedOpToPredOp(N, ISD::UADDSAT, DAG, true); 14046 case Intrinsic::aarch64_sve_uqsub: 14047 return convertMergedOpToPredOp(N, ISD::USUBSAT, DAG, true); 14048 case Intrinsic::aarch64_sve_sqadd_x: 14049 return DAG.getNode(ISD::SADDSAT, SDLoc(N), N->getValueType(0), 14050 N->getOperand(1), N->getOperand(2)); 14051 case Intrinsic::aarch64_sve_sqsub_x: 14052 return DAG.getNode(ISD::SSUBSAT, SDLoc(N), N->getValueType(0), 14053 N->getOperand(1), N->getOperand(2)); 14054 case Intrinsic::aarch64_sve_uqadd_x: 14055 return DAG.getNode(ISD::UADDSAT, SDLoc(N), N->getValueType(0), 14056 N->getOperand(1), N->getOperand(2)); 14057 case Intrinsic::aarch64_sve_uqsub_x: 14058 return DAG.getNode(ISD::USUBSAT, SDLoc(N), N->getValueType(0), 14059 N->getOperand(1), N->getOperand(2)); 14060 case Intrinsic::aarch64_sve_cmphs: 14061 if (!N->getOperand(2).getValueType().isFloatingPoint()) 14062 return DAG.getNode(AArch64ISD::SETCC_MERGE_ZERO, SDLoc(N), 14063 N->getValueType(0), N->getOperand(1), N->getOperand(2), 14064 N->getOperand(3), DAG.getCondCode(ISD::SETUGE)); 14065 break; 14066 case Intrinsic::aarch64_sve_cmphi: 14067 if (!N->getOperand(2).getValueType().isFloatingPoint()) 14068 return DAG.getNode(AArch64ISD::SETCC_MERGE_ZERO, SDLoc(N), 14069 N->getValueType(0), N->getOperand(1), N->getOperand(2), 14070 N->getOperand(3), DAG.getCondCode(ISD::SETUGT)); 14071 break; 14072 case Intrinsic::aarch64_sve_cmpge: 14073 if (!N->getOperand(2).getValueType().isFloatingPoint()) 14074 return DAG.getNode(AArch64ISD::SETCC_MERGE_ZERO, SDLoc(N), 14075 N->getValueType(0), N->getOperand(1), N->getOperand(2), 14076 N->getOperand(3), DAG.getCondCode(ISD::SETGE)); 14077 break; 14078 case Intrinsic::aarch64_sve_cmpgt: 14079 if (!N->getOperand(2).getValueType().isFloatingPoint()) 14080 return DAG.getNode(AArch64ISD::SETCC_MERGE_ZERO, SDLoc(N), 14081 N->getValueType(0), N->getOperand(1), N->getOperand(2), 14082 N->getOperand(3), DAG.getCondCode(ISD::SETGT)); 14083 break; 14084 case Intrinsic::aarch64_sve_cmpeq: 14085 if (!N->getOperand(2).getValueType().isFloatingPoint()) 14086 return DAG.getNode(AArch64ISD::SETCC_MERGE_ZERO, SDLoc(N), 14087 N->getValueType(0), N->getOperand(1), N->getOperand(2), 14088 N->getOperand(3), DAG.getCondCode(ISD::SETEQ)); 14089 break; 14090 case Intrinsic::aarch64_sve_cmpne: 14091 if (!N->getOperand(2).getValueType().isFloatingPoint()) 14092 return DAG.getNode(AArch64ISD::SETCC_MERGE_ZERO, SDLoc(N), 14093 N->getValueType(0), N->getOperand(1), N->getOperand(2), 14094 N->getOperand(3), DAG.getCondCode(ISD::SETNE)); 14095 break; 14096 case Intrinsic::aarch64_sve_fadda: 14097 return combineSVEReductionOrderedFP(N, AArch64ISD::FADDA_PRED, DAG); 14098 case Intrinsic::aarch64_sve_faddv: 14099 return combineSVEReductionFP(N, AArch64ISD::FADDV_PRED, DAG); 14100 case Intrinsic::aarch64_sve_fmaxnmv: 14101 return combineSVEReductionFP(N, AArch64ISD::FMAXNMV_PRED, DAG); 14102 case Intrinsic::aarch64_sve_fmaxv: 14103 return combineSVEReductionFP(N, AArch64ISD::FMAXV_PRED, DAG); 14104 case Intrinsic::aarch64_sve_fminnmv: 14105 return combineSVEReductionFP(N, AArch64ISD::FMINNMV_PRED, DAG); 14106 case Intrinsic::aarch64_sve_fminv: 14107 return combineSVEReductionFP(N, AArch64ISD::FMINV_PRED, DAG); 14108 case Intrinsic::aarch64_sve_sel: 14109 return DAG.getNode(ISD::VSELECT, SDLoc(N), N->getValueType(0), 14110 N->getOperand(1), N->getOperand(2), N->getOperand(3)); 14111 case Intrinsic::aarch64_sve_cmpeq_wide: 14112 return tryConvertSVEWideCompare(N, ISD::SETEQ, DCI, DAG); 14113 case Intrinsic::aarch64_sve_cmpne_wide: 14114 return tryConvertSVEWideCompare(N, ISD::SETNE, DCI, DAG); 14115 case Intrinsic::aarch64_sve_cmpge_wide: 14116 return tryConvertSVEWideCompare(N, ISD::SETGE, DCI, DAG); 14117 case Intrinsic::aarch64_sve_cmpgt_wide: 14118 return tryConvertSVEWideCompare(N, ISD::SETGT, DCI, DAG); 14119 case Intrinsic::aarch64_sve_cmplt_wide: 14120 return tryConvertSVEWideCompare(N, ISD::SETLT, DCI, DAG); 14121 case Intrinsic::aarch64_sve_cmple_wide: 14122 return tryConvertSVEWideCompare(N, ISD::SETLE, DCI, DAG); 14123 case Intrinsic::aarch64_sve_cmphs_wide: 14124 return tryConvertSVEWideCompare(N, ISD::SETUGE, DCI, DAG); 14125 case Intrinsic::aarch64_sve_cmphi_wide: 14126 return tryConvertSVEWideCompare(N, ISD::SETUGT, DCI, DAG); 14127 case Intrinsic::aarch64_sve_cmplo_wide: 14128 return tryConvertSVEWideCompare(N, ISD::SETULT, DCI, DAG); 14129 case Intrinsic::aarch64_sve_cmpls_wide: 14130 return tryConvertSVEWideCompare(N, ISD::SETULE, DCI, DAG); 14131 case Intrinsic::aarch64_sve_ptest_any: 14132 return getPTest(DAG, N->getValueType(0), N->getOperand(1), N->getOperand(2), 14133 AArch64CC::ANY_ACTIVE); 14134 case Intrinsic::aarch64_sve_ptest_first: 14135 return getPTest(DAG, N->getValueType(0), N->getOperand(1), N->getOperand(2), 14136 AArch64CC::FIRST_ACTIVE); 14137 case Intrinsic::aarch64_sve_ptest_last: 14138 return getPTest(DAG, N->getValueType(0), N->getOperand(1), N->getOperand(2), 14139 AArch64CC::LAST_ACTIVE); 14140 } 14141 return SDValue(); 14142 } 14143 14144 static SDValue performExtendCombine(SDNode *N, 14145 TargetLowering::DAGCombinerInfo &DCI, 14146 SelectionDAG &DAG) { 14147 // If we see something like (zext (sabd (extract_high ...), (DUP ...))) then 14148 // we can convert that DUP into another extract_high (of a bigger DUP), which 14149 // helps the backend to decide that an sabdl2 would be useful, saving a real 14150 // extract_high operation. 14151 if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND && 14152 (N->getOperand(0).getOpcode() == AArch64ISD::UABD || 14153 N->getOperand(0).getOpcode() == AArch64ISD::SABD)) { 14154 SDNode *ABDNode = N->getOperand(0).getNode(); 14155 SDValue NewABD = 14156 tryCombineLongOpWithDup(Intrinsic::not_intrinsic, ABDNode, DCI, DAG); 14157 if (!NewABD.getNode()) 14158 return SDValue(); 14159 14160 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0), NewABD); 14161 } 14162 return SDValue(); 14163 } 14164 14165 static SDValue splitStoreSplat(SelectionDAG &DAG, StoreSDNode &St, 14166 SDValue SplatVal, unsigned NumVecElts) { 14167 assert(!St.isTruncatingStore() && "cannot split truncating vector store"); 14168 unsigned OrigAlignment = St.getAlignment(); 14169 unsigned EltOffset = SplatVal.getValueType().getSizeInBits() / 8; 14170 14171 // Create scalar stores. This is at least as good as the code sequence for a 14172 // split unaligned store which is a dup.s, ext.b, and two stores. 14173 // Most of the time the three stores should be replaced by store pair 14174 // instructions (stp). 14175 SDLoc DL(&St); 14176 SDValue BasePtr = St.getBasePtr(); 14177 uint64_t BaseOffset = 0; 14178 14179 const MachinePointerInfo &PtrInfo = St.getPointerInfo(); 14180 SDValue NewST1 = 14181 DAG.getStore(St.getChain(), DL, SplatVal, BasePtr, PtrInfo, 14182 OrigAlignment, St.getMemOperand()->getFlags()); 14183 14184 // As this in ISel, we will not merge this add which may degrade results. 14185 if (BasePtr->getOpcode() == ISD::ADD && 14186 isa<ConstantSDNode>(BasePtr->getOperand(1))) { 14187 BaseOffset = cast<ConstantSDNode>(BasePtr->getOperand(1))->getSExtValue(); 14188 BasePtr = BasePtr->getOperand(0); 14189 } 14190 14191 unsigned Offset = EltOffset; 14192 while (--NumVecElts) { 14193 unsigned Alignment = MinAlign(OrigAlignment, Offset); 14194 SDValue OffsetPtr = 14195 DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr, 14196 DAG.getConstant(BaseOffset + Offset, DL, MVT::i64)); 14197 NewST1 = DAG.getStore(NewST1.getValue(0), DL, SplatVal, OffsetPtr, 14198 PtrInfo.getWithOffset(Offset), Alignment, 14199 St.getMemOperand()->getFlags()); 14200 Offset += EltOffset; 14201 } 14202 return NewST1; 14203 } 14204 14205 // Returns an SVE type that ContentTy can be trivially sign or zero extended 14206 // into. 14207 static MVT getSVEContainerType(EVT ContentTy) { 14208 assert(ContentTy.isSimple() && "No SVE containers for extended types"); 14209 14210 switch (ContentTy.getSimpleVT().SimpleTy) { 14211 default: 14212 llvm_unreachable("No known SVE container for this MVT type"); 14213 case MVT::nxv2i8: 14214 case MVT::nxv2i16: 14215 case MVT::nxv2i32: 14216 case MVT::nxv2i64: 14217 case MVT::nxv2f32: 14218 case MVT::nxv2f64: 14219 return MVT::nxv2i64; 14220 case MVT::nxv4i8: 14221 case MVT::nxv4i16: 14222 case MVT::nxv4i32: 14223 case MVT::nxv4f32: 14224 return MVT::nxv4i32; 14225 case MVT::nxv8i8: 14226 case MVT::nxv8i16: 14227 case MVT::nxv8f16: 14228 case MVT::nxv8bf16: 14229 return MVT::nxv8i16; 14230 case MVT::nxv16i8: 14231 return MVT::nxv16i8; 14232 } 14233 } 14234 14235 static SDValue performLD1Combine(SDNode *N, SelectionDAG &DAG, unsigned Opc) { 14236 SDLoc DL(N); 14237 EVT VT = N->getValueType(0); 14238 14239 if (VT.getSizeInBits().getKnownMinSize() > AArch64::SVEBitsPerBlock) 14240 return SDValue(); 14241 14242 EVT ContainerVT = VT; 14243 if (ContainerVT.isInteger()) 14244 ContainerVT = getSVEContainerType(ContainerVT); 14245 14246 SDVTList VTs = DAG.getVTList(ContainerVT, MVT::Other); 14247 SDValue Ops[] = { N->getOperand(0), // Chain 14248 N->getOperand(2), // Pg 14249 N->getOperand(3), // Base 14250 DAG.getValueType(VT) }; 14251 14252 SDValue Load = DAG.getNode(Opc, DL, VTs, Ops); 14253 SDValue LoadChain = SDValue(Load.getNode(), 1); 14254 14255 if (ContainerVT.isInteger() && (VT != ContainerVT)) 14256 Load = DAG.getNode(ISD::TRUNCATE, DL, VT, Load.getValue(0)); 14257 14258 return DAG.getMergeValues({ Load, LoadChain }, DL); 14259 } 14260 14261 static SDValue performLDNT1Combine(SDNode *N, SelectionDAG &DAG) { 14262 SDLoc DL(N); 14263 EVT VT = N->getValueType(0); 14264 EVT PtrTy = N->getOperand(3).getValueType(); 14265 14266 if (VT == MVT::nxv8bf16 && 14267 !static_cast<const AArch64Subtarget &>(DAG.getSubtarget()).hasBF16()) 14268 return SDValue(); 14269 14270 EVT LoadVT = VT; 14271 if (VT.isFloatingPoint()) 14272 LoadVT = VT.changeTypeToInteger(); 14273 14274 auto *MINode = cast<MemIntrinsicSDNode>(N); 14275 SDValue PassThru = DAG.getConstant(0, DL, LoadVT); 14276 SDValue L = DAG.getMaskedLoad(LoadVT, DL, MINode->getChain(), 14277 MINode->getOperand(3), DAG.getUNDEF(PtrTy), 14278 MINode->getOperand(2), PassThru, 14279 MINode->getMemoryVT(), MINode->getMemOperand(), 14280 ISD::UNINDEXED, ISD::NON_EXTLOAD, false); 14281 14282 if (VT.isFloatingPoint()) { 14283 SDValue Ops[] = { DAG.getNode(ISD::BITCAST, DL, VT, L), L.getValue(1) }; 14284 return DAG.getMergeValues(Ops, DL); 14285 } 14286 14287 return L; 14288 } 14289 14290 template <unsigned Opcode> 14291 static SDValue performLD1ReplicateCombine(SDNode *N, SelectionDAG &DAG) { 14292 static_assert(Opcode == AArch64ISD::LD1RQ_MERGE_ZERO || 14293 Opcode == AArch64ISD::LD1RO_MERGE_ZERO, 14294 "Unsupported opcode."); 14295 SDLoc DL(N); 14296 EVT VT = N->getValueType(0); 14297 if (VT == MVT::nxv8bf16 && 14298 !static_cast<const AArch64Subtarget &>(DAG.getSubtarget()).hasBF16()) 14299 return SDValue(); 14300 14301 EVT LoadVT = VT; 14302 if (VT.isFloatingPoint()) 14303 LoadVT = VT.changeTypeToInteger(); 14304 14305 SDValue Ops[] = {N->getOperand(0), N->getOperand(2), N->getOperand(3)}; 14306 SDValue Load = DAG.getNode(Opcode, DL, {LoadVT, MVT::Other}, Ops); 14307 SDValue LoadChain = SDValue(Load.getNode(), 1); 14308 14309 if (VT.isFloatingPoint()) 14310 Load = DAG.getNode(ISD::BITCAST, DL, VT, Load.getValue(0)); 14311 14312 return DAG.getMergeValues({Load, LoadChain}, DL); 14313 } 14314 14315 static SDValue performST1Combine(SDNode *N, SelectionDAG &DAG) { 14316 SDLoc DL(N); 14317 SDValue Data = N->getOperand(2); 14318 EVT DataVT = Data.getValueType(); 14319 EVT HwSrcVt = getSVEContainerType(DataVT); 14320 SDValue InputVT = DAG.getValueType(DataVT); 14321 14322 if (DataVT == MVT::nxv8bf16 && 14323 !static_cast<const AArch64Subtarget &>(DAG.getSubtarget()).hasBF16()) 14324 return SDValue(); 14325 14326 if (DataVT.isFloatingPoint()) 14327 InputVT = DAG.getValueType(HwSrcVt); 14328 14329 SDValue SrcNew; 14330 if (Data.getValueType().isFloatingPoint()) 14331 SrcNew = DAG.getNode(ISD::BITCAST, DL, HwSrcVt, Data); 14332 else 14333 SrcNew = DAG.getNode(ISD::ANY_EXTEND, DL, HwSrcVt, Data); 14334 14335 SDValue Ops[] = { N->getOperand(0), // Chain 14336 SrcNew, 14337 N->getOperand(4), // Base 14338 N->getOperand(3), // Pg 14339 InputVT 14340 }; 14341 14342 return DAG.getNode(AArch64ISD::ST1_PRED, DL, N->getValueType(0), Ops); 14343 } 14344 14345 static SDValue performSTNT1Combine(SDNode *N, SelectionDAG &DAG) { 14346 SDLoc DL(N); 14347 14348 SDValue Data = N->getOperand(2); 14349 EVT DataVT = Data.getValueType(); 14350 EVT PtrTy = N->getOperand(4).getValueType(); 14351 14352 if (DataVT == MVT::nxv8bf16 && 14353 !static_cast<const AArch64Subtarget &>(DAG.getSubtarget()).hasBF16()) 14354 return SDValue(); 14355 14356 if (DataVT.isFloatingPoint()) 14357 Data = DAG.getNode(ISD::BITCAST, DL, DataVT.changeTypeToInteger(), Data); 14358 14359 auto *MINode = cast<MemIntrinsicSDNode>(N); 14360 return DAG.getMaskedStore(MINode->getChain(), DL, Data, MINode->getOperand(4), 14361 DAG.getUNDEF(PtrTy), MINode->getOperand(3), 14362 MINode->getMemoryVT(), MINode->getMemOperand(), 14363 ISD::UNINDEXED, false, false); 14364 } 14365 14366 /// Replace a splat of zeros to a vector store by scalar stores of WZR/XZR. The 14367 /// load store optimizer pass will merge them to store pair stores. This should 14368 /// be better than a movi to create the vector zero followed by a vector store 14369 /// if the zero constant is not re-used, since one instructions and one register 14370 /// live range will be removed. 14371 /// 14372 /// For example, the final generated code should be: 14373 /// 14374 /// stp xzr, xzr, [x0] 14375 /// 14376 /// instead of: 14377 /// 14378 /// movi v0.2d, #0 14379 /// str q0, [x0] 14380 /// 14381 static SDValue replaceZeroVectorStore(SelectionDAG &DAG, StoreSDNode &St) { 14382 SDValue StVal = St.getValue(); 14383 EVT VT = StVal.getValueType(); 14384 14385 // Avoid scalarizing zero splat stores for scalable vectors. 14386 if (VT.isScalableVector()) 14387 return SDValue(); 14388 14389 // It is beneficial to scalarize a zero splat store for 2 or 3 i64 elements or 14390 // 2, 3 or 4 i32 elements. 14391 int NumVecElts = VT.getVectorNumElements(); 14392 if (!(((NumVecElts == 2 || NumVecElts == 3) && 14393 VT.getVectorElementType().getSizeInBits() == 64) || 14394 ((NumVecElts == 2 || NumVecElts == 3 || NumVecElts == 4) && 14395 VT.getVectorElementType().getSizeInBits() == 32))) 14396 return SDValue(); 14397 14398 if (StVal.getOpcode() != ISD::BUILD_VECTOR) 14399 return SDValue(); 14400 14401 // If the zero constant has more than one use then the vector store could be 14402 // better since the constant mov will be amortized and stp q instructions 14403 // should be able to be formed. 14404 if (!StVal.hasOneUse()) 14405 return SDValue(); 14406 14407 // If the store is truncating then it's going down to i16 or smaller, which 14408 // means it can be implemented in a single store anyway. 14409 if (St.isTruncatingStore()) 14410 return SDValue(); 14411 14412 // If the immediate offset of the address operand is too large for the stp 14413 // instruction, then bail out. 14414 if (DAG.isBaseWithConstantOffset(St.getBasePtr())) { 14415 int64_t Offset = St.getBasePtr()->getConstantOperandVal(1); 14416 if (Offset < -512 || Offset > 504) 14417 return SDValue(); 14418 } 14419 14420 for (int I = 0; I < NumVecElts; ++I) { 14421 SDValue EltVal = StVal.getOperand(I); 14422 if (!isNullConstant(EltVal) && !isNullFPConstant(EltVal)) 14423 return SDValue(); 14424 } 14425 14426 // Use a CopyFromReg WZR/XZR here to prevent 14427 // DAGCombiner::MergeConsecutiveStores from undoing this transformation. 14428 SDLoc DL(&St); 14429 unsigned ZeroReg; 14430 EVT ZeroVT; 14431 if (VT.getVectorElementType().getSizeInBits() == 32) { 14432 ZeroReg = AArch64::WZR; 14433 ZeroVT = MVT::i32; 14434 } else { 14435 ZeroReg = AArch64::XZR; 14436 ZeroVT = MVT::i64; 14437 } 14438 SDValue SplatVal = 14439 DAG.getCopyFromReg(DAG.getEntryNode(), DL, ZeroReg, ZeroVT); 14440 return splitStoreSplat(DAG, St, SplatVal, NumVecElts); 14441 } 14442 14443 /// Replace a splat of a scalar to a vector store by scalar stores of the scalar 14444 /// value. The load store optimizer pass will merge them to store pair stores. 14445 /// This has better performance than a splat of the scalar followed by a split 14446 /// vector store. Even if the stores are not merged it is four stores vs a dup, 14447 /// followed by an ext.b and two stores. 14448 static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode &St) { 14449 SDValue StVal = St.getValue(); 14450 EVT VT = StVal.getValueType(); 14451 14452 // Don't replace floating point stores, they possibly won't be transformed to 14453 // stp because of the store pair suppress pass. 14454 if (VT.isFloatingPoint()) 14455 return SDValue(); 14456 14457 // We can express a splat as store pair(s) for 2 or 4 elements. 14458 unsigned NumVecElts = VT.getVectorNumElements(); 14459 if (NumVecElts != 4 && NumVecElts != 2) 14460 return SDValue(); 14461 14462 // If the store is truncating then it's going down to i16 or smaller, which 14463 // means it can be implemented in a single store anyway. 14464 if (St.isTruncatingStore()) 14465 return SDValue(); 14466 14467 // Check that this is a splat. 14468 // Make sure that each of the relevant vector element locations are inserted 14469 // to, i.e. 0 and 1 for v2i64 and 0, 1, 2, 3 for v4i32. 14470 std::bitset<4> IndexNotInserted((1 << NumVecElts) - 1); 14471 SDValue SplatVal; 14472 for (unsigned I = 0; I < NumVecElts; ++I) { 14473 // Check for insert vector elements. 14474 if (StVal.getOpcode() != ISD::INSERT_VECTOR_ELT) 14475 return SDValue(); 14476 14477 // Check that same value is inserted at each vector element. 14478 if (I == 0) 14479 SplatVal = StVal.getOperand(1); 14480 else if (StVal.getOperand(1) != SplatVal) 14481 return SDValue(); 14482 14483 // Check insert element index. 14484 ConstantSDNode *CIndex = dyn_cast<ConstantSDNode>(StVal.getOperand(2)); 14485 if (!CIndex) 14486 return SDValue(); 14487 uint64_t IndexVal = CIndex->getZExtValue(); 14488 if (IndexVal >= NumVecElts) 14489 return SDValue(); 14490 IndexNotInserted.reset(IndexVal); 14491 14492 StVal = StVal.getOperand(0); 14493 } 14494 // Check that all vector element locations were inserted to. 14495 if (IndexNotInserted.any()) 14496 return SDValue(); 14497 14498 return splitStoreSplat(DAG, St, SplatVal, NumVecElts); 14499 } 14500 14501 static SDValue splitStores(SDNode *N, TargetLowering::DAGCombinerInfo &DCI, 14502 SelectionDAG &DAG, 14503 const AArch64Subtarget *Subtarget) { 14504 14505 StoreSDNode *S = cast<StoreSDNode>(N); 14506 if (S->isVolatile() || S->isIndexed()) 14507 return SDValue(); 14508 14509 SDValue StVal = S->getValue(); 14510 EVT VT = StVal.getValueType(); 14511 14512 if (!VT.isFixedLengthVector()) 14513 return SDValue(); 14514 14515 // If we get a splat of zeros, convert this vector store to a store of 14516 // scalars. They will be merged into store pairs of xzr thereby removing one 14517 // instruction and one register. 14518 if (SDValue ReplacedZeroSplat = replaceZeroVectorStore(DAG, *S)) 14519 return ReplacedZeroSplat; 14520 14521 // FIXME: The logic for deciding if an unaligned store should be split should 14522 // be included in TLI.allowsMisalignedMemoryAccesses(), and there should be 14523 // a call to that function here. 14524 14525 if (!Subtarget->isMisaligned128StoreSlow()) 14526 return SDValue(); 14527 14528 // Don't split at -Oz. 14529 if (DAG.getMachineFunction().getFunction().hasMinSize()) 14530 return SDValue(); 14531 14532 // Don't split v2i64 vectors. Memcpy lowering produces those and splitting 14533 // those up regresses performance on micro-benchmarks and olden/bh. 14534 if (VT.getVectorNumElements() < 2 || VT == MVT::v2i64) 14535 return SDValue(); 14536 14537 // Split unaligned 16B stores. They are terrible for performance. 14538 // Don't split stores with alignment of 1 or 2. Code that uses clang vector 14539 // extensions can use this to mark that it does not want splitting to happen 14540 // (by underspecifying alignment to be 1 or 2). Furthermore, the chance of 14541 // eliminating alignment hazards is only 1 in 8 for alignment of 2. 14542 if (VT.getSizeInBits() != 128 || S->getAlignment() >= 16 || 14543 S->getAlignment() <= 2) 14544 return SDValue(); 14545 14546 // If we get a splat of a scalar convert this vector store to a store of 14547 // scalars. They will be merged into store pairs thereby removing two 14548 // instructions. 14549 if (SDValue ReplacedSplat = replaceSplatVectorStore(DAG, *S)) 14550 return ReplacedSplat; 14551 14552 SDLoc DL(S); 14553 14554 // Split VT into two. 14555 EVT HalfVT = VT.getHalfNumVectorElementsVT(*DAG.getContext()); 14556 unsigned NumElts = HalfVT.getVectorNumElements(); 14557 SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal, 14558 DAG.getConstant(0, DL, MVT::i64)); 14559 SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal, 14560 DAG.getConstant(NumElts, DL, MVT::i64)); 14561 SDValue BasePtr = S->getBasePtr(); 14562 SDValue NewST1 = 14563 DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(), 14564 S->getAlignment(), S->getMemOperand()->getFlags()); 14565 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr, 14566 DAG.getConstant(8, DL, MVT::i64)); 14567 return DAG.getStore(NewST1.getValue(0), DL, SubVector1, OffsetPtr, 14568 S->getPointerInfo(), S->getAlignment(), 14569 S->getMemOperand()->getFlags()); 14570 } 14571 14572 static SDValue performUzpCombine(SDNode *N, SelectionDAG &DAG) { 14573 SDLoc DL(N); 14574 SDValue Op0 = N->getOperand(0); 14575 SDValue Op1 = N->getOperand(1); 14576 EVT ResVT = N->getValueType(0); 14577 14578 // uzp1(unpklo(uzp1(x, y)), z) => uzp1(x, z) 14579 if (Op0.getOpcode() == AArch64ISD::UUNPKLO) { 14580 if (Op0.getOperand(0).getOpcode() == AArch64ISD::UZP1) { 14581 SDValue X = Op0.getOperand(0).getOperand(0); 14582 return DAG.getNode(AArch64ISD::UZP1, DL, ResVT, X, Op1); 14583 } 14584 } 14585 14586 // uzp1(x, unpkhi(uzp1(y, z))) => uzp1(x, z) 14587 if (Op1.getOpcode() == AArch64ISD::UUNPKHI) { 14588 if (Op1.getOperand(0).getOpcode() == AArch64ISD::UZP1) { 14589 SDValue Z = Op1.getOperand(0).getOperand(1); 14590 return DAG.getNode(AArch64ISD::UZP1, DL, ResVT, Op0, Z); 14591 } 14592 } 14593 14594 return SDValue(); 14595 } 14596 14597 static SDValue performGLD1Combine(SDNode *N, SelectionDAG &DAG) { 14598 unsigned Opc = N->getOpcode(); 14599 14600 assert(((Opc >= AArch64ISD::GLD1_MERGE_ZERO && // unsigned gather loads 14601 Opc <= AArch64ISD::GLD1_IMM_MERGE_ZERO) || 14602 (Opc >= AArch64ISD::GLD1S_MERGE_ZERO && // signed gather loads 14603 Opc <= AArch64ISD::GLD1S_IMM_MERGE_ZERO)) && 14604 "Invalid opcode."); 14605 14606 const bool Scaled = Opc == AArch64ISD::GLD1_SCALED_MERGE_ZERO || 14607 Opc == AArch64ISD::GLD1S_SCALED_MERGE_ZERO; 14608 const bool Signed = Opc == AArch64ISD::GLD1S_MERGE_ZERO || 14609 Opc == AArch64ISD::GLD1S_SCALED_MERGE_ZERO; 14610 const bool Extended = Opc == AArch64ISD::GLD1_SXTW_MERGE_ZERO || 14611 Opc == AArch64ISD::GLD1_SXTW_SCALED_MERGE_ZERO || 14612 Opc == AArch64ISD::GLD1_UXTW_MERGE_ZERO || 14613 Opc == AArch64ISD::GLD1_UXTW_SCALED_MERGE_ZERO; 14614 14615 SDLoc DL(N); 14616 SDValue Chain = N->getOperand(0); 14617 SDValue Pg = N->getOperand(1); 14618 SDValue Base = N->getOperand(2); 14619 SDValue Offset = N->getOperand(3); 14620 SDValue Ty = N->getOperand(4); 14621 14622 EVT ResVT = N->getValueType(0); 14623 14624 const auto OffsetOpc = Offset.getOpcode(); 14625 const bool OffsetIsZExt = 14626 OffsetOpc == AArch64ISD::ZERO_EXTEND_INREG_MERGE_PASSTHRU; 14627 const bool OffsetIsSExt = 14628 OffsetOpc == AArch64ISD::SIGN_EXTEND_INREG_MERGE_PASSTHRU; 14629 14630 // Fold sign/zero extensions of vector offsets into GLD1 nodes where possible. 14631 if (!Extended && (OffsetIsSExt || OffsetIsZExt)) { 14632 SDValue ExtPg = Offset.getOperand(0); 14633 VTSDNode *ExtFrom = cast<VTSDNode>(Offset.getOperand(2).getNode()); 14634 EVT ExtFromEVT = ExtFrom->getVT().getVectorElementType(); 14635 14636 // If the predicate for the sign- or zero-extended offset is the 14637 // same as the predicate used for this load and the sign-/zero-extension 14638 // was from a 32-bits... 14639 if (ExtPg == Pg && ExtFromEVT == MVT::i32) { 14640 SDValue UnextendedOffset = Offset.getOperand(1); 14641 14642 unsigned NewOpc = getGatherVecOpcode(Scaled, OffsetIsSExt, true); 14643 if (Signed) 14644 NewOpc = getSignExtendedGatherOpcode(NewOpc); 14645 14646 return DAG.getNode(NewOpc, DL, {ResVT, MVT::Other}, 14647 {Chain, Pg, Base, UnextendedOffset, Ty}); 14648 } 14649 } 14650 14651 return SDValue(); 14652 } 14653 14654 /// Optimize a vector shift instruction and its operand if shifted out 14655 /// bits are not used. 14656 static SDValue performVectorShiftCombine(SDNode *N, 14657 const AArch64TargetLowering &TLI, 14658 TargetLowering::DAGCombinerInfo &DCI) { 14659 assert(N->getOpcode() == AArch64ISD::VASHR || 14660 N->getOpcode() == AArch64ISD::VLSHR); 14661 14662 SDValue Op = N->getOperand(0); 14663 unsigned OpScalarSize = Op.getScalarValueSizeInBits(); 14664 14665 unsigned ShiftImm = N->getConstantOperandVal(1); 14666 assert(OpScalarSize > ShiftImm && "Invalid shift imm"); 14667 14668 APInt ShiftedOutBits = APInt::getLowBitsSet(OpScalarSize, ShiftImm); 14669 APInt DemandedMask = ~ShiftedOutBits; 14670 14671 if (TLI.SimplifyDemandedBits(Op, DemandedMask, DCI)) 14672 return SDValue(N, 0); 14673 14674 return SDValue(); 14675 } 14676 14677 /// Target-specific DAG combine function for post-increment LD1 (lane) and 14678 /// post-increment LD1R. 14679 static SDValue performPostLD1Combine(SDNode *N, 14680 TargetLowering::DAGCombinerInfo &DCI, 14681 bool IsLaneOp) { 14682 if (DCI.isBeforeLegalizeOps()) 14683 return SDValue(); 14684 14685 SelectionDAG &DAG = DCI.DAG; 14686 EVT VT = N->getValueType(0); 14687 14688 if (VT.isScalableVector()) 14689 return SDValue(); 14690 14691 unsigned LoadIdx = IsLaneOp ? 1 : 0; 14692 SDNode *LD = N->getOperand(LoadIdx).getNode(); 14693 // If it is not LOAD, can not do such combine. 14694 if (LD->getOpcode() != ISD::LOAD) 14695 return SDValue(); 14696 14697 // The vector lane must be a constant in the LD1LANE opcode. 14698 SDValue Lane; 14699 if (IsLaneOp) { 14700 Lane = N->getOperand(2); 14701 auto *LaneC = dyn_cast<ConstantSDNode>(Lane); 14702 if (!LaneC || LaneC->getZExtValue() >= VT.getVectorNumElements()) 14703 return SDValue(); 14704 } 14705 14706 LoadSDNode *LoadSDN = cast<LoadSDNode>(LD); 14707 EVT MemVT = LoadSDN->getMemoryVT(); 14708 // Check if memory operand is the same type as the vector element. 14709 if (MemVT != VT.getVectorElementType()) 14710 return SDValue(); 14711 14712 // Check if there are other uses. If so, do not combine as it will introduce 14713 // an extra load. 14714 for (SDNode::use_iterator UI = LD->use_begin(), UE = LD->use_end(); UI != UE; 14715 ++UI) { 14716 if (UI.getUse().getResNo() == 1) // Ignore uses of the chain result. 14717 continue; 14718 if (*UI != N) 14719 return SDValue(); 14720 } 14721 14722 SDValue Addr = LD->getOperand(1); 14723 SDValue Vector = N->getOperand(0); 14724 // Search for a use of the address operand that is an increment. 14725 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(), UE = 14726 Addr.getNode()->use_end(); UI != UE; ++UI) { 14727 SDNode *User = *UI; 14728 if (User->getOpcode() != ISD::ADD 14729 || UI.getUse().getResNo() != Addr.getResNo()) 14730 continue; 14731 14732 // If the increment is a constant, it must match the memory ref size. 14733 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0); 14734 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) { 14735 uint32_t IncVal = CInc->getZExtValue(); 14736 unsigned NumBytes = VT.getScalarSizeInBits() / 8; 14737 if (IncVal != NumBytes) 14738 continue; 14739 Inc = DAG.getRegister(AArch64::XZR, MVT::i64); 14740 } 14741 14742 // To avoid cycle construction make sure that neither the load nor the add 14743 // are predecessors to each other or the Vector. 14744 SmallPtrSet<const SDNode *, 32> Visited; 14745 SmallVector<const SDNode *, 16> Worklist; 14746 Visited.insert(Addr.getNode()); 14747 Worklist.push_back(User); 14748 Worklist.push_back(LD); 14749 Worklist.push_back(Vector.getNode()); 14750 if (SDNode::hasPredecessorHelper(LD, Visited, Worklist) || 14751 SDNode::hasPredecessorHelper(User, Visited, Worklist)) 14752 continue; 14753 14754 SmallVector<SDValue, 8> Ops; 14755 Ops.push_back(LD->getOperand(0)); // Chain 14756 if (IsLaneOp) { 14757 Ops.push_back(Vector); // The vector to be inserted 14758 Ops.push_back(Lane); // The lane to be inserted in the vector 14759 } 14760 Ops.push_back(Addr); 14761 Ops.push_back(Inc); 14762 14763 EVT Tys[3] = { VT, MVT::i64, MVT::Other }; 14764 SDVTList SDTys = DAG.getVTList(Tys); 14765 unsigned NewOp = IsLaneOp ? AArch64ISD::LD1LANEpost : AArch64ISD::LD1DUPpost; 14766 SDValue UpdN = DAG.getMemIntrinsicNode(NewOp, SDLoc(N), SDTys, Ops, 14767 MemVT, 14768 LoadSDN->getMemOperand()); 14769 14770 // Update the uses. 14771 SDValue NewResults[] = { 14772 SDValue(LD, 0), // The result of load 14773 SDValue(UpdN.getNode(), 2) // Chain 14774 }; 14775 DCI.CombineTo(LD, NewResults); 14776 DCI.CombineTo(N, SDValue(UpdN.getNode(), 0)); // Dup/Inserted Result 14777 DCI.CombineTo(User, SDValue(UpdN.getNode(), 1)); // Write back register 14778 14779 break; 14780 } 14781 return SDValue(); 14782 } 14783 14784 /// Simplify ``Addr`` given that the top byte of it is ignored by HW during 14785 /// address translation. 14786 static bool performTBISimplification(SDValue Addr, 14787 TargetLowering::DAGCombinerInfo &DCI, 14788 SelectionDAG &DAG) { 14789 APInt DemandedMask = APInt::getLowBitsSet(64, 56); 14790 KnownBits Known; 14791 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(), 14792 !DCI.isBeforeLegalizeOps()); 14793 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 14794 if (TLI.SimplifyDemandedBits(Addr, DemandedMask, Known, TLO)) { 14795 DCI.CommitTargetLoweringOpt(TLO); 14796 return true; 14797 } 14798 return false; 14799 } 14800 14801 static SDValue performSTORECombine(SDNode *N, 14802 TargetLowering::DAGCombinerInfo &DCI, 14803 SelectionDAG &DAG, 14804 const AArch64Subtarget *Subtarget) { 14805 if (SDValue Split = splitStores(N, DCI, DAG, Subtarget)) 14806 return Split; 14807 14808 if (Subtarget->supportsAddressTopByteIgnored() && 14809 performTBISimplification(N->getOperand(2), DCI, DAG)) 14810 return SDValue(N, 0); 14811 14812 return SDValue(); 14813 } 14814 14815 /// Target-specific DAG combine function for NEON load/store intrinsics 14816 /// to merge base address updates. 14817 static SDValue performNEONPostLDSTCombine(SDNode *N, 14818 TargetLowering::DAGCombinerInfo &DCI, 14819 SelectionDAG &DAG) { 14820 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer()) 14821 return SDValue(); 14822 14823 unsigned AddrOpIdx = N->getNumOperands() - 1; 14824 SDValue Addr = N->getOperand(AddrOpIdx); 14825 14826 // Search for a use of the address operand that is an increment. 14827 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(), 14828 UE = Addr.getNode()->use_end(); UI != UE; ++UI) { 14829 SDNode *User = *UI; 14830 if (User->getOpcode() != ISD::ADD || 14831 UI.getUse().getResNo() != Addr.getResNo()) 14832 continue; 14833 14834 // Check that the add is independent of the load/store. Otherwise, folding 14835 // it would create a cycle. 14836 SmallPtrSet<const SDNode *, 32> Visited; 14837 SmallVector<const SDNode *, 16> Worklist; 14838 Visited.insert(Addr.getNode()); 14839 Worklist.push_back(N); 14840 Worklist.push_back(User); 14841 if (SDNode::hasPredecessorHelper(N, Visited, Worklist) || 14842 SDNode::hasPredecessorHelper(User, Visited, Worklist)) 14843 continue; 14844 14845 // Find the new opcode for the updating load/store. 14846 bool IsStore = false; 14847 bool IsLaneOp = false; 14848 bool IsDupOp = false; 14849 unsigned NewOpc = 0; 14850 unsigned NumVecs = 0; 14851 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue(); 14852 switch (IntNo) { 14853 default: llvm_unreachable("unexpected intrinsic for Neon base update"); 14854 case Intrinsic::aarch64_neon_ld2: NewOpc = AArch64ISD::LD2post; 14855 NumVecs = 2; break; 14856 case Intrinsic::aarch64_neon_ld3: NewOpc = AArch64ISD::LD3post; 14857 NumVecs = 3; break; 14858 case Intrinsic::aarch64_neon_ld4: NewOpc = AArch64ISD::LD4post; 14859 NumVecs = 4; break; 14860 case Intrinsic::aarch64_neon_st2: NewOpc = AArch64ISD::ST2post; 14861 NumVecs = 2; IsStore = true; break; 14862 case Intrinsic::aarch64_neon_st3: NewOpc = AArch64ISD::ST3post; 14863 NumVecs = 3; IsStore = true; break; 14864 case Intrinsic::aarch64_neon_st4: NewOpc = AArch64ISD::ST4post; 14865 NumVecs = 4; IsStore = true; break; 14866 case Intrinsic::aarch64_neon_ld1x2: NewOpc = AArch64ISD::LD1x2post; 14867 NumVecs = 2; break; 14868 case Intrinsic::aarch64_neon_ld1x3: NewOpc = AArch64ISD::LD1x3post; 14869 NumVecs = 3; break; 14870 case Intrinsic::aarch64_neon_ld1x4: NewOpc = AArch64ISD::LD1x4post; 14871 NumVecs = 4; break; 14872 case Intrinsic::aarch64_neon_st1x2: NewOpc = AArch64ISD::ST1x2post; 14873 NumVecs = 2; IsStore = true; break; 14874 case Intrinsic::aarch64_neon_st1x3: NewOpc = AArch64ISD::ST1x3post; 14875 NumVecs = 3; IsStore = true; break; 14876 case Intrinsic::aarch64_neon_st1x4: NewOpc = AArch64ISD::ST1x4post; 14877 NumVecs = 4; IsStore = true; break; 14878 case Intrinsic::aarch64_neon_ld2r: NewOpc = AArch64ISD::LD2DUPpost; 14879 NumVecs = 2; IsDupOp = true; break; 14880 case Intrinsic::aarch64_neon_ld3r: NewOpc = AArch64ISD::LD3DUPpost; 14881 NumVecs = 3; IsDupOp = true; break; 14882 case Intrinsic::aarch64_neon_ld4r: NewOpc = AArch64ISD::LD4DUPpost; 14883 NumVecs = 4; IsDupOp = true; break; 14884 case Intrinsic::aarch64_neon_ld2lane: NewOpc = AArch64ISD::LD2LANEpost; 14885 NumVecs = 2; IsLaneOp = true; break; 14886 case Intrinsic::aarch64_neon_ld3lane: NewOpc = AArch64ISD::LD3LANEpost; 14887 NumVecs = 3; IsLaneOp = true; break; 14888 case Intrinsic::aarch64_neon_ld4lane: NewOpc = AArch64ISD::LD4LANEpost; 14889 NumVecs = 4; IsLaneOp = true; break; 14890 case Intrinsic::aarch64_neon_st2lane: NewOpc = AArch64ISD::ST2LANEpost; 14891 NumVecs = 2; IsStore = true; IsLaneOp = true; break; 14892 case Intrinsic::aarch64_neon_st3lane: NewOpc = AArch64ISD::ST3LANEpost; 14893 NumVecs = 3; IsStore = true; IsLaneOp = true; break; 14894 case Intrinsic::aarch64_neon_st4lane: NewOpc = AArch64ISD::ST4LANEpost; 14895 NumVecs = 4; IsStore = true; IsLaneOp = true; break; 14896 } 14897 14898 EVT VecTy; 14899 if (IsStore) 14900 VecTy = N->getOperand(2).getValueType(); 14901 else 14902 VecTy = N->getValueType(0); 14903 14904 // If the increment is a constant, it must match the memory ref size. 14905 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0); 14906 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) { 14907 uint32_t IncVal = CInc->getZExtValue(); 14908 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8; 14909 if (IsLaneOp || IsDupOp) 14910 NumBytes /= VecTy.getVectorNumElements(); 14911 if (IncVal != NumBytes) 14912 continue; 14913 Inc = DAG.getRegister(AArch64::XZR, MVT::i64); 14914 } 14915 SmallVector<SDValue, 8> Ops; 14916 Ops.push_back(N->getOperand(0)); // Incoming chain 14917 // Load lane and store have vector list as input. 14918 if (IsLaneOp || IsStore) 14919 for (unsigned i = 2; i < AddrOpIdx; ++i) 14920 Ops.push_back(N->getOperand(i)); 14921 Ops.push_back(Addr); // Base register 14922 Ops.push_back(Inc); 14923 14924 // Return Types. 14925 EVT Tys[6]; 14926 unsigned NumResultVecs = (IsStore ? 0 : NumVecs); 14927 unsigned n; 14928 for (n = 0; n < NumResultVecs; ++n) 14929 Tys[n] = VecTy; 14930 Tys[n++] = MVT::i64; // Type of write back register 14931 Tys[n] = MVT::Other; // Type of the chain 14932 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs + 2)); 14933 14934 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N); 14935 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys, Ops, 14936 MemInt->getMemoryVT(), 14937 MemInt->getMemOperand()); 14938 14939 // Update the uses. 14940 std::vector<SDValue> NewResults; 14941 for (unsigned i = 0; i < NumResultVecs; ++i) { 14942 NewResults.push_back(SDValue(UpdN.getNode(), i)); 14943 } 14944 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs + 1)); 14945 DCI.CombineTo(N, NewResults); 14946 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs)); 14947 14948 break; 14949 } 14950 return SDValue(); 14951 } 14952 14953 // Checks to see if the value is the prescribed width and returns information 14954 // about its extension mode. 14955 static 14956 bool checkValueWidth(SDValue V, unsigned width, ISD::LoadExtType &ExtType) { 14957 ExtType = ISD::NON_EXTLOAD; 14958 switch(V.getNode()->getOpcode()) { 14959 default: 14960 return false; 14961 case ISD::LOAD: { 14962 LoadSDNode *LoadNode = cast<LoadSDNode>(V.getNode()); 14963 if ((LoadNode->getMemoryVT() == MVT::i8 && width == 8) 14964 || (LoadNode->getMemoryVT() == MVT::i16 && width == 16)) { 14965 ExtType = LoadNode->getExtensionType(); 14966 return true; 14967 } 14968 return false; 14969 } 14970 case ISD::AssertSext: { 14971 VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1)); 14972 if ((TypeNode->getVT() == MVT::i8 && width == 8) 14973 || (TypeNode->getVT() == MVT::i16 && width == 16)) { 14974 ExtType = ISD::SEXTLOAD; 14975 return true; 14976 } 14977 return false; 14978 } 14979 case ISD::AssertZext: { 14980 VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1)); 14981 if ((TypeNode->getVT() == MVT::i8 && width == 8) 14982 || (TypeNode->getVT() == MVT::i16 && width == 16)) { 14983 ExtType = ISD::ZEXTLOAD; 14984 return true; 14985 } 14986 return false; 14987 } 14988 case ISD::Constant: 14989 case ISD::TargetConstant: { 14990 return std::abs(cast<ConstantSDNode>(V.getNode())->getSExtValue()) < 14991 1LL << (width - 1); 14992 } 14993 } 14994 14995 return true; 14996 } 14997 14998 // This function does a whole lot of voodoo to determine if the tests are 14999 // equivalent without and with a mask. Essentially what happens is that given a 15000 // DAG resembling: 15001 // 15002 // +-------------+ +-------------+ +-------------+ +-------------+ 15003 // | Input | | AddConstant | | CompConstant| | CC | 15004 // +-------------+ +-------------+ +-------------+ +-------------+ 15005 // | | | | 15006 // V V | +----------+ 15007 // +-------------+ +----+ | | 15008 // | ADD | |0xff| | | 15009 // +-------------+ +----+ | | 15010 // | | | | 15011 // V V | | 15012 // +-------------+ | | 15013 // | AND | | | 15014 // +-------------+ | | 15015 // | | | 15016 // +-----+ | | 15017 // | | | 15018 // V V V 15019 // +-------------+ 15020 // | CMP | 15021 // +-------------+ 15022 // 15023 // The AND node may be safely removed for some combinations of inputs. In 15024 // particular we need to take into account the extension type of the Input, 15025 // the exact values of AddConstant, CompConstant, and CC, along with the nominal 15026 // width of the input (this can work for any width inputs, the above graph is 15027 // specific to 8 bits. 15028 // 15029 // The specific equations were worked out by generating output tables for each 15030 // AArch64CC value in terms of and AddConstant (w1), CompConstant(w2). The 15031 // problem was simplified by working with 4 bit inputs, which means we only 15032 // needed to reason about 24 distinct bit patterns: 8 patterns unique to zero 15033 // extension (8,15), 8 patterns unique to sign extensions (-8,-1), and 8 15034 // patterns present in both extensions (0,7). For every distinct set of 15035 // AddConstant and CompConstants bit patterns we can consider the masked and 15036 // unmasked versions to be equivalent if the result of this function is true for 15037 // all 16 distinct bit patterns of for the current extension type of Input (w0). 15038 // 15039 // sub w8, w0, w1 15040 // and w10, w8, #0x0f 15041 // cmp w8, w2 15042 // cset w9, AArch64CC 15043 // cmp w10, w2 15044 // cset w11, AArch64CC 15045 // cmp w9, w11 15046 // cset w0, eq 15047 // ret 15048 // 15049 // Since the above function shows when the outputs are equivalent it defines 15050 // when it is safe to remove the AND. Unfortunately it only runs on AArch64 and 15051 // would be expensive to run during compiles. The equations below were written 15052 // in a test harness that confirmed they gave equivalent outputs to the above 15053 // for all inputs function, so they can be used determine if the removal is 15054 // legal instead. 15055 // 15056 // isEquivalentMaskless() is the code for testing if the AND can be removed 15057 // factored out of the DAG recognition as the DAG can take several forms. 15058 15059 static bool isEquivalentMaskless(unsigned CC, unsigned width, 15060 ISD::LoadExtType ExtType, int AddConstant, 15061 int CompConstant) { 15062 // By being careful about our equations and only writing the in term 15063 // symbolic values and well known constants (0, 1, -1, MaxUInt) we can 15064 // make them generally applicable to all bit widths. 15065 int MaxUInt = (1 << width); 15066 15067 // For the purposes of these comparisons sign extending the type is 15068 // equivalent to zero extending the add and displacing it by half the integer 15069 // width. Provided we are careful and make sure our equations are valid over 15070 // the whole range we can just adjust the input and avoid writing equations 15071 // for sign extended inputs. 15072 if (ExtType == ISD::SEXTLOAD) 15073 AddConstant -= (1 << (width-1)); 15074 15075 switch(CC) { 15076 case AArch64CC::LE: 15077 case AArch64CC::GT: 15078 if ((AddConstant == 0) || 15079 (CompConstant == MaxUInt - 1 && AddConstant < 0) || 15080 (AddConstant >= 0 && CompConstant < 0) || 15081 (AddConstant <= 0 && CompConstant <= 0 && CompConstant < AddConstant)) 15082 return true; 15083 break; 15084 case AArch64CC::LT: 15085 case AArch64CC::GE: 15086 if ((AddConstant == 0) || 15087 (AddConstant >= 0 && CompConstant <= 0) || 15088 (AddConstant <= 0 && CompConstant <= 0 && CompConstant <= AddConstant)) 15089 return true; 15090 break; 15091 case AArch64CC::HI: 15092 case AArch64CC::LS: 15093 if ((AddConstant >= 0 && CompConstant < 0) || 15094 (AddConstant <= 0 && CompConstant >= -1 && 15095 CompConstant < AddConstant + MaxUInt)) 15096 return true; 15097 break; 15098 case AArch64CC::PL: 15099 case AArch64CC::MI: 15100 if ((AddConstant == 0) || 15101 (AddConstant > 0 && CompConstant <= 0) || 15102 (AddConstant < 0 && CompConstant <= AddConstant)) 15103 return true; 15104 break; 15105 case AArch64CC::LO: 15106 case AArch64CC::HS: 15107 if ((AddConstant >= 0 && CompConstant <= 0) || 15108 (AddConstant <= 0 && CompConstant >= 0 && 15109 CompConstant <= AddConstant + MaxUInt)) 15110 return true; 15111 break; 15112 case AArch64CC::EQ: 15113 case AArch64CC::NE: 15114 if ((AddConstant > 0 && CompConstant < 0) || 15115 (AddConstant < 0 && CompConstant >= 0 && 15116 CompConstant < AddConstant + MaxUInt) || 15117 (AddConstant >= 0 && CompConstant >= 0 && 15118 CompConstant >= AddConstant) || 15119 (AddConstant <= 0 && CompConstant < 0 && CompConstant < AddConstant)) 15120 return true; 15121 break; 15122 case AArch64CC::VS: 15123 case AArch64CC::VC: 15124 case AArch64CC::AL: 15125 case AArch64CC::NV: 15126 return true; 15127 case AArch64CC::Invalid: 15128 break; 15129 } 15130 15131 return false; 15132 } 15133 15134 static 15135 SDValue performCONDCombine(SDNode *N, 15136 TargetLowering::DAGCombinerInfo &DCI, 15137 SelectionDAG &DAG, unsigned CCIndex, 15138 unsigned CmpIndex) { 15139 unsigned CC = cast<ConstantSDNode>(N->getOperand(CCIndex))->getSExtValue(); 15140 SDNode *SubsNode = N->getOperand(CmpIndex).getNode(); 15141 unsigned CondOpcode = SubsNode->getOpcode(); 15142 15143 if (CondOpcode != AArch64ISD::SUBS) 15144 return SDValue(); 15145 15146 // There is a SUBS feeding this condition. Is it fed by a mask we can 15147 // use? 15148 15149 SDNode *AndNode = SubsNode->getOperand(0).getNode(); 15150 unsigned MaskBits = 0; 15151 15152 if (AndNode->getOpcode() != ISD::AND) 15153 return SDValue(); 15154 15155 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(AndNode->getOperand(1))) { 15156 uint32_t CNV = CN->getZExtValue(); 15157 if (CNV == 255) 15158 MaskBits = 8; 15159 else if (CNV == 65535) 15160 MaskBits = 16; 15161 } 15162 15163 if (!MaskBits) 15164 return SDValue(); 15165 15166 SDValue AddValue = AndNode->getOperand(0); 15167 15168 if (AddValue.getOpcode() != ISD::ADD) 15169 return SDValue(); 15170 15171 // The basic dag structure is correct, grab the inputs and validate them. 15172 15173 SDValue AddInputValue1 = AddValue.getNode()->getOperand(0); 15174 SDValue AddInputValue2 = AddValue.getNode()->getOperand(1); 15175 SDValue SubsInputValue = SubsNode->getOperand(1); 15176 15177 // The mask is present and the provenance of all the values is a smaller type, 15178 // lets see if the mask is superfluous. 15179 15180 if (!isa<ConstantSDNode>(AddInputValue2.getNode()) || 15181 !isa<ConstantSDNode>(SubsInputValue.getNode())) 15182 return SDValue(); 15183 15184 ISD::LoadExtType ExtType; 15185 15186 if (!checkValueWidth(SubsInputValue, MaskBits, ExtType) || 15187 !checkValueWidth(AddInputValue2, MaskBits, ExtType) || 15188 !checkValueWidth(AddInputValue1, MaskBits, ExtType) ) 15189 return SDValue(); 15190 15191 if(!isEquivalentMaskless(CC, MaskBits, ExtType, 15192 cast<ConstantSDNode>(AddInputValue2.getNode())->getSExtValue(), 15193 cast<ConstantSDNode>(SubsInputValue.getNode())->getSExtValue())) 15194 return SDValue(); 15195 15196 // The AND is not necessary, remove it. 15197 15198 SDVTList VTs = DAG.getVTList(SubsNode->getValueType(0), 15199 SubsNode->getValueType(1)); 15200 SDValue Ops[] = { AddValue, SubsNode->getOperand(1) }; 15201 15202 SDValue NewValue = DAG.getNode(CondOpcode, SDLoc(SubsNode), VTs, Ops); 15203 DAG.ReplaceAllUsesWith(SubsNode, NewValue.getNode()); 15204 15205 return SDValue(N, 0); 15206 } 15207 15208 // Optimize compare with zero and branch. 15209 static SDValue performBRCONDCombine(SDNode *N, 15210 TargetLowering::DAGCombinerInfo &DCI, 15211 SelectionDAG &DAG) { 15212 MachineFunction &MF = DAG.getMachineFunction(); 15213 // Speculation tracking/SLH assumes that optimized TB(N)Z/CB(N)Z instructions 15214 // will not be produced, as they are conditional branch instructions that do 15215 // not set flags. 15216 if (MF.getFunction().hasFnAttribute(Attribute::SpeculativeLoadHardening)) 15217 return SDValue(); 15218 15219 if (SDValue NV = performCONDCombine(N, DCI, DAG, 2, 3)) 15220 N = NV.getNode(); 15221 SDValue Chain = N->getOperand(0); 15222 SDValue Dest = N->getOperand(1); 15223 SDValue CCVal = N->getOperand(2); 15224 SDValue Cmp = N->getOperand(3); 15225 15226 assert(isa<ConstantSDNode>(CCVal) && "Expected a ConstantSDNode here!"); 15227 unsigned CC = cast<ConstantSDNode>(CCVal)->getZExtValue(); 15228 if (CC != AArch64CC::EQ && CC != AArch64CC::NE) 15229 return SDValue(); 15230 15231 unsigned CmpOpc = Cmp.getOpcode(); 15232 if (CmpOpc != AArch64ISD::ADDS && CmpOpc != AArch64ISD::SUBS) 15233 return SDValue(); 15234 15235 // Only attempt folding if there is only one use of the flag and no use of the 15236 // value. 15237 if (!Cmp->hasNUsesOfValue(0, 0) || !Cmp->hasNUsesOfValue(1, 1)) 15238 return SDValue(); 15239 15240 SDValue LHS = Cmp.getOperand(0); 15241 SDValue RHS = Cmp.getOperand(1); 15242 15243 assert(LHS.getValueType() == RHS.getValueType() && 15244 "Expected the value type to be the same for both operands!"); 15245 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64) 15246 return SDValue(); 15247 15248 if (isNullConstant(LHS)) 15249 std::swap(LHS, RHS); 15250 15251 if (!isNullConstant(RHS)) 15252 return SDValue(); 15253 15254 if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA || 15255 LHS.getOpcode() == ISD::SRL) 15256 return SDValue(); 15257 15258 // Fold the compare into the branch instruction. 15259 SDValue BR; 15260 if (CC == AArch64CC::EQ) 15261 BR = DAG.getNode(AArch64ISD::CBZ, SDLoc(N), MVT::Other, Chain, LHS, Dest); 15262 else 15263 BR = DAG.getNode(AArch64ISD::CBNZ, SDLoc(N), MVT::Other, Chain, LHS, Dest); 15264 15265 // Do not add new nodes to DAG combiner worklist. 15266 DCI.CombineTo(N, BR, false); 15267 15268 return SDValue(); 15269 } 15270 15271 // Optimize CSEL instructions 15272 static SDValue performCSELCombine(SDNode *N, 15273 TargetLowering::DAGCombinerInfo &DCI, 15274 SelectionDAG &DAG) { 15275 // CSEL x, x, cc -> x 15276 if (N->getOperand(0) == N->getOperand(1)) 15277 return N->getOperand(0); 15278 15279 return performCONDCombine(N, DCI, DAG, 2, 3); 15280 } 15281 15282 // Optimize some simple tbz/tbnz cases. Returns the new operand and bit to test 15283 // as well as whether the test should be inverted. This code is required to 15284 // catch these cases (as opposed to standard dag combines) because 15285 // AArch64ISD::TBZ is matched during legalization. 15286 static SDValue getTestBitOperand(SDValue Op, unsigned &Bit, bool &Invert, 15287 SelectionDAG &DAG) { 15288 15289 if (!Op->hasOneUse()) 15290 return Op; 15291 15292 // We don't handle undef/constant-fold cases below, as they should have 15293 // already been taken care of (e.g. and of 0, test of undefined shifted bits, 15294 // etc.) 15295 15296 // (tbz (trunc x), b) -> (tbz x, b) 15297 // This case is just here to enable more of the below cases to be caught. 15298 if (Op->getOpcode() == ISD::TRUNCATE && 15299 Bit < Op->getValueType(0).getSizeInBits()) { 15300 return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG); 15301 } 15302 15303 // (tbz (any_ext x), b) -> (tbz x, b) if we don't use the extended bits. 15304 if (Op->getOpcode() == ISD::ANY_EXTEND && 15305 Bit < Op->getOperand(0).getValueSizeInBits()) { 15306 return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG); 15307 } 15308 15309 if (Op->getNumOperands() != 2) 15310 return Op; 15311 15312 auto *C = dyn_cast<ConstantSDNode>(Op->getOperand(1)); 15313 if (!C) 15314 return Op; 15315 15316 switch (Op->getOpcode()) { 15317 default: 15318 return Op; 15319 15320 // (tbz (and x, m), b) -> (tbz x, b) 15321 case ISD::AND: 15322 if ((C->getZExtValue() >> Bit) & 1) 15323 return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG); 15324 return Op; 15325 15326 // (tbz (shl x, c), b) -> (tbz x, b-c) 15327 case ISD::SHL: 15328 if (C->getZExtValue() <= Bit && 15329 (Bit - C->getZExtValue()) < Op->getValueType(0).getSizeInBits()) { 15330 Bit = Bit - C->getZExtValue(); 15331 return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG); 15332 } 15333 return Op; 15334 15335 // (tbz (sra x, c), b) -> (tbz x, b+c) or (tbz x, msb) if b+c is > # bits in x 15336 case ISD::SRA: 15337 Bit = Bit + C->getZExtValue(); 15338 if (Bit >= Op->getValueType(0).getSizeInBits()) 15339 Bit = Op->getValueType(0).getSizeInBits() - 1; 15340 return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG); 15341 15342 // (tbz (srl x, c), b) -> (tbz x, b+c) 15343 case ISD::SRL: 15344 if ((Bit + C->getZExtValue()) < Op->getValueType(0).getSizeInBits()) { 15345 Bit = Bit + C->getZExtValue(); 15346 return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG); 15347 } 15348 return Op; 15349 15350 // (tbz (xor x, -1), b) -> (tbnz x, b) 15351 case ISD::XOR: 15352 if ((C->getZExtValue() >> Bit) & 1) 15353 Invert = !Invert; 15354 return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG); 15355 } 15356 } 15357 15358 // Optimize test single bit zero/non-zero and branch. 15359 static SDValue performTBZCombine(SDNode *N, 15360 TargetLowering::DAGCombinerInfo &DCI, 15361 SelectionDAG &DAG) { 15362 unsigned Bit = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue(); 15363 bool Invert = false; 15364 SDValue TestSrc = N->getOperand(1); 15365 SDValue NewTestSrc = getTestBitOperand(TestSrc, Bit, Invert, DAG); 15366 15367 if (TestSrc == NewTestSrc) 15368 return SDValue(); 15369 15370 unsigned NewOpc = N->getOpcode(); 15371 if (Invert) { 15372 if (NewOpc == AArch64ISD::TBZ) 15373 NewOpc = AArch64ISD::TBNZ; 15374 else { 15375 assert(NewOpc == AArch64ISD::TBNZ); 15376 NewOpc = AArch64ISD::TBZ; 15377 } 15378 } 15379 15380 SDLoc DL(N); 15381 return DAG.getNode(NewOpc, DL, MVT::Other, N->getOperand(0), NewTestSrc, 15382 DAG.getConstant(Bit, DL, MVT::i64), N->getOperand(3)); 15383 } 15384 15385 // vselect (v1i1 setcc) -> 15386 // vselect (v1iXX setcc) (XX is the size of the compared operand type) 15387 // FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as 15388 // condition. If it can legalize "VSELECT v1i1" correctly, no need to combine 15389 // such VSELECT. 15390 static SDValue performVSelectCombine(SDNode *N, SelectionDAG &DAG) { 15391 SDValue N0 = N->getOperand(0); 15392 EVT CCVT = N0.getValueType(); 15393 15394 // Check for sign pattern (VSELECT setgt, iN lhs, -1, 1, -1) and transform 15395 // into (OR (ASR lhs, N-1), 1), which requires less instructions for the 15396 // supported types. 15397 SDValue SetCC = N->getOperand(0); 15398 if (SetCC.getOpcode() == ISD::SETCC && 15399 SetCC.getOperand(2) == DAG.getCondCode(ISD::SETGT)) { 15400 SDValue CmpLHS = SetCC.getOperand(0); 15401 EVT VT = CmpLHS.getValueType(); 15402 SDNode *CmpRHS = SetCC.getOperand(1).getNode(); 15403 SDNode *SplatLHS = N->getOperand(1).getNode(); 15404 SDNode *SplatRHS = N->getOperand(2).getNode(); 15405 APInt SplatLHSVal; 15406 if (CmpLHS.getValueType() == N->getOperand(1).getValueType() && 15407 VT.isSimple() && 15408 is_contained( 15409 makeArrayRef({MVT::v8i8, MVT::v16i8, MVT::v4i16, MVT::v8i16, 15410 MVT::v2i32, MVT::v4i32, MVT::v2i64}), 15411 VT.getSimpleVT().SimpleTy) && 15412 ISD::isConstantSplatVector(SplatLHS, SplatLHSVal) && 15413 SplatLHSVal.isOneValue() && ISD::isConstantSplatVectorAllOnes(CmpRHS) && 15414 ISD::isConstantSplatVectorAllOnes(SplatRHS)) { 15415 unsigned NumElts = VT.getVectorNumElements(); 15416 SmallVector<SDValue, 8> Ops( 15417 NumElts, DAG.getConstant(VT.getScalarSizeInBits() - 1, SDLoc(N), 15418 VT.getScalarType())); 15419 SDValue Val = DAG.getBuildVector(VT, SDLoc(N), Ops); 15420 15421 auto Shift = DAG.getNode(ISD::SRA, SDLoc(N), VT, CmpLHS, Val); 15422 auto Or = DAG.getNode(ISD::OR, SDLoc(N), VT, Shift, N->getOperand(1)); 15423 return Or; 15424 } 15425 } 15426 15427 if (N0.getOpcode() != ISD::SETCC || 15428 CCVT.getVectorElementCount() != ElementCount::getFixed(1) || 15429 CCVT.getVectorElementType() != MVT::i1) 15430 return SDValue(); 15431 15432 EVT ResVT = N->getValueType(0); 15433 EVT CmpVT = N0.getOperand(0).getValueType(); 15434 // Only combine when the result type is of the same size as the compared 15435 // operands. 15436 if (ResVT.getSizeInBits() != CmpVT.getSizeInBits()) 15437 return SDValue(); 15438 15439 SDValue IfTrue = N->getOperand(1); 15440 SDValue IfFalse = N->getOperand(2); 15441 SetCC = DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(), 15442 N0.getOperand(0), N0.getOperand(1), 15443 cast<CondCodeSDNode>(N0.getOperand(2))->get()); 15444 return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC, 15445 IfTrue, IfFalse); 15446 } 15447 15448 /// A vector select: "(select vL, vR, (setcc LHS, RHS))" is best performed with 15449 /// the compare-mask instructions rather than going via NZCV, even if LHS and 15450 /// RHS are really scalar. This replaces any scalar setcc in the above pattern 15451 /// with a vector one followed by a DUP shuffle on the result. 15452 static SDValue performSelectCombine(SDNode *N, 15453 TargetLowering::DAGCombinerInfo &DCI) { 15454 SelectionDAG &DAG = DCI.DAG; 15455 SDValue N0 = N->getOperand(0); 15456 EVT ResVT = N->getValueType(0); 15457 15458 if (N0.getOpcode() != ISD::SETCC) 15459 return SDValue(); 15460 15461 if (ResVT.isScalableVector()) 15462 return SDValue(); 15463 15464 // Make sure the SETCC result is either i1 (initial DAG), or i32, the lowered 15465 // scalar SetCCResultType. We also don't expect vectors, because we assume 15466 // that selects fed by vector SETCCs are canonicalized to VSELECT. 15467 assert((N0.getValueType() == MVT::i1 || N0.getValueType() == MVT::i32) && 15468 "Scalar-SETCC feeding SELECT has unexpected result type!"); 15469 15470 // If NumMaskElts == 0, the comparison is larger than select result. The 15471 // largest real NEON comparison is 64-bits per lane, which means the result is 15472 // at most 32-bits and an illegal vector. Just bail out for now. 15473 EVT SrcVT = N0.getOperand(0).getValueType(); 15474 15475 // Don't try to do this optimization when the setcc itself has i1 operands. 15476 // There are no legal vectors of i1, so this would be pointless. 15477 if (SrcVT == MVT::i1) 15478 return SDValue(); 15479 15480 int NumMaskElts = ResVT.getSizeInBits() / SrcVT.getSizeInBits(); 15481 if (!ResVT.isVector() || NumMaskElts == 0) 15482 return SDValue(); 15483 15484 SrcVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumMaskElts); 15485 EVT CCVT = SrcVT.changeVectorElementTypeToInteger(); 15486 15487 // Also bail out if the vector CCVT isn't the same size as ResVT. 15488 // This can happen if the SETCC operand size doesn't divide the ResVT size 15489 // (e.g., f64 vs v3f32). 15490 if (CCVT.getSizeInBits() != ResVT.getSizeInBits()) 15491 return SDValue(); 15492 15493 // Make sure we didn't create illegal types, if we're not supposed to. 15494 assert(DCI.isBeforeLegalize() || 15495 DAG.getTargetLoweringInfo().isTypeLegal(SrcVT)); 15496 15497 // First perform a vector comparison, where lane 0 is the one we're interested 15498 // in. 15499 SDLoc DL(N0); 15500 SDValue LHS = 15501 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(0)); 15502 SDValue RHS = 15503 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(1)); 15504 SDValue SetCC = DAG.getNode(ISD::SETCC, DL, CCVT, LHS, RHS, N0.getOperand(2)); 15505 15506 // Now duplicate the comparison mask we want across all other lanes. 15507 SmallVector<int, 8> DUPMask(CCVT.getVectorNumElements(), 0); 15508 SDValue Mask = DAG.getVectorShuffle(CCVT, DL, SetCC, SetCC, DUPMask); 15509 Mask = DAG.getNode(ISD::BITCAST, DL, 15510 ResVT.changeVectorElementTypeToInteger(), Mask); 15511 15512 return DAG.getSelect(DL, ResVT, Mask, N->getOperand(1), N->getOperand(2)); 15513 } 15514 15515 /// Get rid of unnecessary NVCASTs (that don't change the type). 15516 static SDValue performNVCASTCombine(SDNode *N) { 15517 if (N->getValueType(0) == N->getOperand(0).getValueType()) 15518 return N->getOperand(0); 15519 15520 return SDValue(); 15521 } 15522 15523 // If all users of the globaladdr are of the form (globaladdr + constant), find 15524 // the smallest constant, fold it into the globaladdr's offset and rewrite the 15525 // globaladdr as (globaladdr + constant) - constant. 15526 static SDValue performGlobalAddressCombine(SDNode *N, SelectionDAG &DAG, 15527 const AArch64Subtarget *Subtarget, 15528 const TargetMachine &TM) { 15529 auto *GN = cast<GlobalAddressSDNode>(N); 15530 if (Subtarget->ClassifyGlobalReference(GN->getGlobal(), TM) != 15531 AArch64II::MO_NO_FLAG) 15532 return SDValue(); 15533 15534 uint64_t MinOffset = -1ull; 15535 for (SDNode *N : GN->uses()) { 15536 if (N->getOpcode() != ISD::ADD) 15537 return SDValue(); 15538 auto *C = dyn_cast<ConstantSDNode>(N->getOperand(0)); 15539 if (!C) 15540 C = dyn_cast<ConstantSDNode>(N->getOperand(1)); 15541 if (!C) 15542 return SDValue(); 15543 MinOffset = std::min(MinOffset, C->getZExtValue()); 15544 } 15545 uint64_t Offset = MinOffset + GN->getOffset(); 15546 15547 // Require that the new offset is larger than the existing one. Otherwise, we 15548 // can end up oscillating between two possible DAGs, for example, 15549 // (add (add globaladdr + 10, -1), 1) and (add globaladdr + 9, 1). 15550 if (Offset <= uint64_t(GN->getOffset())) 15551 return SDValue(); 15552 15553 // Check whether folding this offset is legal. It must not go out of bounds of 15554 // the referenced object to avoid violating the code model, and must be 15555 // smaller than 2^21 because this is the largest offset expressible in all 15556 // object formats. 15557 // 15558 // This check also prevents us from folding negative offsets, which will end 15559 // up being treated in the same way as large positive ones. They could also 15560 // cause code model violations, and aren't really common enough to matter. 15561 if (Offset >= (1 << 21)) 15562 return SDValue(); 15563 15564 const GlobalValue *GV = GN->getGlobal(); 15565 Type *T = GV->getValueType(); 15566 if (!T->isSized() || 15567 Offset > GV->getParent()->getDataLayout().getTypeAllocSize(T)) 15568 return SDValue(); 15569 15570 SDLoc DL(GN); 15571 SDValue Result = DAG.getGlobalAddress(GV, DL, MVT::i64, Offset); 15572 return DAG.getNode(ISD::SUB, DL, MVT::i64, Result, 15573 DAG.getConstant(MinOffset, DL, MVT::i64)); 15574 } 15575 15576 // Turns the vector of indices into a vector of byte offstes by scaling Offset 15577 // by (BitWidth / 8). 15578 static SDValue getScaledOffsetForBitWidth(SelectionDAG &DAG, SDValue Offset, 15579 SDLoc DL, unsigned BitWidth) { 15580 assert(Offset.getValueType().isScalableVector() && 15581 "This method is only for scalable vectors of offsets"); 15582 15583 SDValue Shift = DAG.getConstant(Log2_32(BitWidth / 8), DL, MVT::i64); 15584 SDValue SplatShift = DAG.getNode(ISD::SPLAT_VECTOR, DL, MVT::nxv2i64, Shift); 15585 15586 return DAG.getNode(ISD::SHL, DL, MVT::nxv2i64, Offset, SplatShift); 15587 } 15588 15589 /// Check if the value of \p OffsetInBytes can be used as an immediate for 15590 /// the gather load/prefetch and scatter store instructions with vector base and 15591 /// immediate offset addressing mode: 15592 /// 15593 /// [<Zn>.[S|D]{, #<imm>}] 15594 /// 15595 /// where <imm> = sizeof(<T>) * k, for k = 0, 1, ..., 31. 15596 15597 inline static bool isValidImmForSVEVecImmAddrMode(unsigned OffsetInBytes, 15598 unsigned ScalarSizeInBytes) { 15599 // The immediate is not a multiple of the scalar size. 15600 if (OffsetInBytes % ScalarSizeInBytes) 15601 return false; 15602 15603 // The immediate is out of range. 15604 if (OffsetInBytes / ScalarSizeInBytes > 31) 15605 return false; 15606 15607 return true; 15608 } 15609 15610 /// Check if the value of \p Offset represents a valid immediate for the SVE 15611 /// gather load/prefetch and scatter store instructiona with vector base and 15612 /// immediate offset addressing mode: 15613 /// 15614 /// [<Zn>.[S|D]{, #<imm>}] 15615 /// 15616 /// where <imm> = sizeof(<T>) * k, for k = 0, 1, ..., 31. 15617 static bool isValidImmForSVEVecImmAddrMode(SDValue Offset, 15618 unsigned ScalarSizeInBytes) { 15619 ConstantSDNode *OffsetConst = dyn_cast<ConstantSDNode>(Offset.getNode()); 15620 return OffsetConst && isValidImmForSVEVecImmAddrMode( 15621 OffsetConst->getZExtValue(), ScalarSizeInBytes); 15622 } 15623 15624 static SDValue performScatterStoreCombine(SDNode *N, SelectionDAG &DAG, 15625 unsigned Opcode, 15626 bool OnlyPackedOffsets = true) { 15627 const SDValue Src = N->getOperand(2); 15628 const EVT SrcVT = Src->getValueType(0); 15629 assert(SrcVT.isScalableVector() && 15630 "Scatter stores are only possible for SVE vectors"); 15631 15632 SDLoc DL(N); 15633 MVT SrcElVT = SrcVT.getVectorElementType().getSimpleVT(); 15634 15635 // Make sure that source data will fit into an SVE register 15636 if (SrcVT.getSizeInBits().getKnownMinSize() > AArch64::SVEBitsPerBlock) 15637 return SDValue(); 15638 15639 // For FPs, ACLE only supports _packed_ single and double precision types. 15640 if (SrcElVT.isFloatingPoint()) 15641 if ((SrcVT != MVT::nxv4f32) && (SrcVT != MVT::nxv2f64)) 15642 return SDValue(); 15643 15644 // Depending on the addressing mode, this is either a pointer or a vector of 15645 // pointers (that fits into one register) 15646 SDValue Base = N->getOperand(4); 15647 // Depending on the addressing mode, this is either a single offset or a 15648 // vector of offsets (that fits into one register) 15649 SDValue Offset = N->getOperand(5); 15650 15651 // For "scalar + vector of indices", just scale the indices. This only 15652 // applies to non-temporal scatters because there's no instruction that takes 15653 // indicies. 15654 if (Opcode == AArch64ISD::SSTNT1_INDEX_PRED) { 15655 Offset = 15656 getScaledOffsetForBitWidth(DAG, Offset, DL, SrcElVT.getSizeInBits()); 15657 Opcode = AArch64ISD::SSTNT1_PRED; 15658 } 15659 15660 // In the case of non-temporal gather loads there's only one SVE instruction 15661 // per data-size: "scalar + vector", i.e. 15662 // * stnt1{b|h|w|d} { z0.s }, p0/z, [z0.s, x0] 15663 // Since we do have intrinsics that allow the arguments to be in a different 15664 // order, we may need to swap them to match the spec. 15665 if (Opcode == AArch64ISD::SSTNT1_PRED && Offset.getValueType().isVector()) 15666 std::swap(Base, Offset); 15667 15668 // SST1_IMM requires that the offset is an immediate that is: 15669 // * a multiple of #SizeInBytes, 15670 // * in the range [0, 31 x #SizeInBytes], 15671 // where #SizeInBytes is the size in bytes of the stored items. For 15672 // immediates outside that range and non-immediate scalar offsets use SST1 or 15673 // SST1_UXTW instead. 15674 if (Opcode == AArch64ISD::SST1_IMM_PRED) { 15675 if (!isValidImmForSVEVecImmAddrMode(Offset, 15676 SrcVT.getScalarSizeInBits() / 8)) { 15677 if (MVT::nxv4i32 == Base.getValueType().getSimpleVT().SimpleTy) 15678 Opcode = AArch64ISD::SST1_UXTW_PRED; 15679 else 15680 Opcode = AArch64ISD::SST1_PRED; 15681 15682 std::swap(Base, Offset); 15683 } 15684 } 15685 15686 auto &TLI = DAG.getTargetLoweringInfo(); 15687 if (!TLI.isTypeLegal(Base.getValueType())) 15688 return SDValue(); 15689 15690 // Some scatter store variants allow unpacked offsets, but only as nxv2i32 15691 // vectors. These are implicitly sign (sxtw) or zero (zxtw) extend to 15692 // nxv2i64. Legalize accordingly. 15693 if (!OnlyPackedOffsets && 15694 Offset.getValueType().getSimpleVT().SimpleTy == MVT::nxv2i32) 15695 Offset = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::nxv2i64, Offset).getValue(0); 15696 15697 if (!TLI.isTypeLegal(Offset.getValueType())) 15698 return SDValue(); 15699 15700 // Source value type that is representable in hardware 15701 EVT HwSrcVt = getSVEContainerType(SrcVT); 15702 15703 // Keep the original type of the input data to store - this is needed to be 15704 // able to select the correct instruction, e.g. ST1B, ST1H, ST1W and ST1D. For 15705 // FP values we want the integer equivalent, so just use HwSrcVt. 15706 SDValue InputVT = DAG.getValueType(SrcVT); 15707 if (SrcVT.isFloatingPoint()) 15708 InputVT = DAG.getValueType(HwSrcVt); 15709 15710 SDVTList VTs = DAG.getVTList(MVT::Other); 15711 SDValue SrcNew; 15712 15713 if (Src.getValueType().isFloatingPoint()) 15714 SrcNew = DAG.getNode(ISD::BITCAST, DL, HwSrcVt, Src); 15715 else 15716 SrcNew = DAG.getNode(ISD::ANY_EXTEND, DL, HwSrcVt, Src); 15717 15718 SDValue Ops[] = {N->getOperand(0), // Chain 15719 SrcNew, 15720 N->getOperand(3), // Pg 15721 Base, 15722 Offset, 15723 InputVT}; 15724 15725 return DAG.getNode(Opcode, DL, VTs, Ops); 15726 } 15727 15728 static SDValue performGatherLoadCombine(SDNode *N, SelectionDAG &DAG, 15729 unsigned Opcode, 15730 bool OnlyPackedOffsets = true) { 15731 const EVT RetVT = N->getValueType(0); 15732 assert(RetVT.isScalableVector() && 15733 "Gather loads are only possible for SVE vectors"); 15734 15735 SDLoc DL(N); 15736 15737 // Make sure that the loaded data will fit into an SVE register 15738 if (RetVT.getSizeInBits().getKnownMinSize() > AArch64::SVEBitsPerBlock) 15739 return SDValue(); 15740 15741 // Depending on the addressing mode, this is either a pointer or a vector of 15742 // pointers (that fits into one register) 15743 SDValue Base = N->getOperand(3); 15744 // Depending on the addressing mode, this is either a single offset or a 15745 // vector of offsets (that fits into one register) 15746 SDValue Offset = N->getOperand(4); 15747 15748 // For "scalar + vector of indices", just scale the indices. This only 15749 // applies to non-temporal gathers because there's no instruction that takes 15750 // indicies. 15751 if (Opcode == AArch64ISD::GLDNT1_INDEX_MERGE_ZERO) { 15752 Offset = getScaledOffsetForBitWidth(DAG, Offset, DL, 15753 RetVT.getScalarSizeInBits()); 15754 Opcode = AArch64ISD::GLDNT1_MERGE_ZERO; 15755 } 15756 15757 // In the case of non-temporal gather loads there's only one SVE instruction 15758 // per data-size: "scalar + vector", i.e. 15759 // * ldnt1{b|h|w|d} { z0.s }, p0/z, [z0.s, x0] 15760 // Since we do have intrinsics that allow the arguments to be in a different 15761 // order, we may need to swap them to match the spec. 15762 if (Opcode == AArch64ISD::GLDNT1_MERGE_ZERO && 15763 Offset.getValueType().isVector()) 15764 std::swap(Base, Offset); 15765 15766 // GLD{FF}1_IMM requires that the offset is an immediate that is: 15767 // * a multiple of #SizeInBytes, 15768 // * in the range [0, 31 x #SizeInBytes], 15769 // where #SizeInBytes is the size in bytes of the loaded items. For 15770 // immediates outside that range and non-immediate scalar offsets use 15771 // GLD1_MERGE_ZERO or GLD1_UXTW_MERGE_ZERO instead. 15772 if (Opcode == AArch64ISD::GLD1_IMM_MERGE_ZERO || 15773 Opcode == AArch64ISD::GLDFF1_IMM_MERGE_ZERO) { 15774 if (!isValidImmForSVEVecImmAddrMode(Offset, 15775 RetVT.getScalarSizeInBits() / 8)) { 15776 if (MVT::nxv4i32 == Base.getValueType().getSimpleVT().SimpleTy) 15777 Opcode = (Opcode == AArch64ISD::GLD1_IMM_MERGE_ZERO) 15778 ? AArch64ISD::GLD1_UXTW_MERGE_ZERO 15779 : AArch64ISD::GLDFF1_UXTW_MERGE_ZERO; 15780 else 15781 Opcode = (Opcode == AArch64ISD::GLD1_IMM_MERGE_ZERO) 15782 ? AArch64ISD::GLD1_MERGE_ZERO 15783 : AArch64ISD::GLDFF1_MERGE_ZERO; 15784 15785 std::swap(Base, Offset); 15786 } 15787 } 15788 15789 auto &TLI = DAG.getTargetLoweringInfo(); 15790 if (!TLI.isTypeLegal(Base.getValueType())) 15791 return SDValue(); 15792 15793 // Some gather load variants allow unpacked offsets, but only as nxv2i32 15794 // vectors. These are implicitly sign (sxtw) or zero (zxtw) extend to 15795 // nxv2i64. Legalize accordingly. 15796 if (!OnlyPackedOffsets && 15797 Offset.getValueType().getSimpleVT().SimpleTy == MVT::nxv2i32) 15798 Offset = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::nxv2i64, Offset).getValue(0); 15799 15800 // Return value type that is representable in hardware 15801 EVT HwRetVt = getSVEContainerType(RetVT); 15802 15803 // Keep the original output value type around - this is needed to be able to 15804 // select the correct instruction, e.g. LD1B, LD1H, LD1W and LD1D. For FP 15805 // values we want the integer equivalent, so just use HwRetVT. 15806 SDValue OutVT = DAG.getValueType(RetVT); 15807 if (RetVT.isFloatingPoint()) 15808 OutVT = DAG.getValueType(HwRetVt); 15809 15810 SDVTList VTs = DAG.getVTList(HwRetVt, MVT::Other); 15811 SDValue Ops[] = {N->getOperand(0), // Chain 15812 N->getOperand(2), // Pg 15813 Base, Offset, OutVT}; 15814 15815 SDValue Load = DAG.getNode(Opcode, DL, VTs, Ops); 15816 SDValue LoadChain = SDValue(Load.getNode(), 1); 15817 15818 if (RetVT.isInteger() && (RetVT != HwRetVt)) 15819 Load = DAG.getNode(ISD::TRUNCATE, DL, RetVT, Load.getValue(0)); 15820 15821 // If the original return value was FP, bitcast accordingly. Doing it here 15822 // means that we can avoid adding TableGen patterns for FPs. 15823 if (RetVT.isFloatingPoint()) 15824 Load = DAG.getNode(ISD::BITCAST, DL, RetVT, Load.getValue(0)); 15825 15826 return DAG.getMergeValues({Load, LoadChain}, DL); 15827 } 15828 15829 static SDValue 15830 performSignExtendInRegCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI, 15831 SelectionDAG &DAG) { 15832 SDLoc DL(N); 15833 SDValue Src = N->getOperand(0); 15834 unsigned Opc = Src->getOpcode(); 15835 15836 // Sign extend of an unsigned unpack -> signed unpack 15837 if (Opc == AArch64ISD::UUNPKHI || Opc == AArch64ISD::UUNPKLO) { 15838 15839 unsigned SOpc = Opc == AArch64ISD::UUNPKHI ? AArch64ISD::SUNPKHI 15840 : AArch64ISD::SUNPKLO; 15841 15842 // Push the sign extend to the operand of the unpack 15843 // This is necessary where, for example, the operand of the unpack 15844 // is another unpack: 15845 // 4i32 sign_extend_inreg (4i32 uunpklo(8i16 uunpklo (16i8 opnd)), from 4i8) 15846 // -> 15847 // 4i32 sunpklo (8i16 sign_extend_inreg(8i16 uunpklo (16i8 opnd), from 8i8) 15848 // -> 15849 // 4i32 sunpklo(8i16 sunpklo(16i8 opnd)) 15850 SDValue ExtOp = Src->getOperand(0); 15851 auto VT = cast<VTSDNode>(N->getOperand(1))->getVT(); 15852 EVT EltTy = VT.getVectorElementType(); 15853 (void)EltTy; 15854 15855 assert((EltTy == MVT::i8 || EltTy == MVT::i16 || EltTy == MVT::i32) && 15856 "Sign extending from an invalid type"); 15857 15858 EVT ExtVT = VT.getDoubleNumVectorElementsVT(*DAG.getContext()); 15859 15860 SDValue Ext = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ExtOp.getValueType(), 15861 ExtOp, DAG.getValueType(ExtVT)); 15862 15863 return DAG.getNode(SOpc, DL, N->getValueType(0), Ext); 15864 } 15865 15866 if (DCI.isBeforeLegalizeOps()) 15867 return SDValue(); 15868 15869 if (!EnableCombineMGatherIntrinsics) 15870 return SDValue(); 15871 15872 // SVE load nodes (e.g. AArch64ISD::GLD1) are straightforward candidates 15873 // for DAG Combine with SIGN_EXTEND_INREG. Bail out for all other nodes. 15874 unsigned NewOpc; 15875 unsigned MemVTOpNum = 4; 15876 switch (Opc) { 15877 case AArch64ISD::LD1_MERGE_ZERO: 15878 NewOpc = AArch64ISD::LD1S_MERGE_ZERO; 15879 MemVTOpNum = 3; 15880 break; 15881 case AArch64ISD::LDNF1_MERGE_ZERO: 15882 NewOpc = AArch64ISD::LDNF1S_MERGE_ZERO; 15883 MemVTOpNum = 3; 15884 break; 15885 case AArch64ISD::LDFF1_MERGE_ZERO: 15886 NewOpc = AArch64ISD::LDFF1S_MERGE_ZERO; 15887 MemVTOpNum = 3; 15888 break; 15889 case AArch64ISD::GLD1_MERGE_ZERO: 15890 NewOpc = AArch64ISD::GLD1S_MERGE_ZERO; 15891 break; 15892 case AArch64ISD::GLD1_SCALED_MERGE_ZERO: 15893 NewOpc = AArch64ISD::GLD1S_SCALED_MERGE_ZERO; 15894 break; 15895 case AArch64ISD::GLD1_SXTW_MERGE_ZERO: 15896 NewOpc = AArch64ISD::GLD1S_SXTW_MERGE_ZERO; 15897 break; 15898 case AArch64ISD::GLD1_SXTW_SCALED_MERGE_ZERO: 15899 NewOpc = AArch64ISD::GLD1S_SXTW_SCALED_MERGE_ZERO; 15900 break; 15901 case AArch64ISD::GLD1_UXTW_MERGE_ZERO: 15902 NewOpc = AArch64ISD::GLD1S_UXTW_MERGE_ZERO; 15903 break; 15904 case AArch64ISD::GLD1_UXTW_SCALED_MERGE_ZERO: 15905 NewOpc = AArch64ISD::GLD1S_UXTW_SCALED_MERGE_ZERO; 15906 break; 15907 case AArch64ISD::GLD1_IMM_MERGE_ZERO: 15908 NewOpc = AArch64ISD::GLD1S_IMM_MERGE_ZERO; 15909 break; 15910 case AArch64ISD::GLDFF1_MERGE_ZERO: 15911 NewOpc = AArch64ISD::GLDFF1S_MERGE_ZERO; 15912 break; 15913 case AArch64ISD::GLDFF1_SCALED_MERGE_ZERO: 15914 NewOpc = AArch64ISD::GLDFF1S_SCALED_MERGE_ZERO; 15915 break; 15916 case AArch64ISD::GLDFF1_SXTW_MERGE_ZERO: 15917 NewOpc = AArch64ISD::GLDFF1S_SXTW_MERGE_ZERO; 15918 break; 15919 case AArch64ISD::GLDFF1_SXTW_SCALED_MERGE_ZERO: 15920 NewOpc = AArch64ISD::GLDFF1S_SXTW_SCALED_MERGE_ZERO; 15921 break; 15922 case AArch64ISD::GLDFF1_UXTW_MERGE_ZERO: 15923 NewOpc = AArch64ISD::GLDFF1S_UXTW_MERGE_ZERO; 15924 break; 15925 case AArch64ISD::GLDFF1_UXTW_SCALED_MERGE_ZERO: 15926 NewOpc = AArch64ISD::GLDFF1S_UXTW_SCALED_MERGE_ZERO; 15927 break; 15928 case AArch64ISD::GLDFF1_IMM_MERGE_ZERO: 15929 NewOpc = AArch64ISD::GLDFF1S_IMM_MERGE_ZERO; 15930 break; 15931 case AArch64ISD::GLDNT1_MERGE_ZERO: 15932 NewOpc = AArch64ISD::GLDNT1S_MERGE_ZERO; 15933 break; 15934 default: 15935 return SDValue(); 15936 } 15937 15938 EVT SignExtSrcVT = cast<VTSDNode>(N->getOperand(1))->getVT(); 15939 EVT SrcMemVT = cast<VTSDNode>(Src->getOperand(MemVTOpNum))->getVT(); 15940 15941 if ((SignExtSrcVT != SrcMemVT) || !Src.hasOneUse()) 15942 return SDValue(); 15943 15944 EVT DstVT = N->getValueType(0); 15945 SDVTList VTs = DAG.getVTList(DstVT, MVT::Other); 15946 15947 SmallVector<SDValue, 5> Ops; 15948 for (unsigned I = 0; I < Src->getNumOperands(); ++I) 15949 Ops.push_back(Src->getOperand(I)); 15950 15951 SDValue ExtLoad = DAG.getNode(NewOpc, SDLoc(N), VTs, Ops); 15952 DCI.CombineTo(N, ExtLoad); 15953 DCI.CombineTo(Src.getNode(), ExtLoad, ExtLoad.getValue(1)); 15954 15955 // Return N so it doesn't get rechecked 15956 return SDValue(N, 0); 15957 } 15958 15959 /// Legalize the gather prefetch (scalar + vector addressing mode) when the 15960 /// offset vector is an unpacked 32-bit scalable vector. The other cases (Offset 15961 /// != nxv2i32) do not need legalization. 15962 static SDValue legalizeSVEGatherPrefetchOffsVec(SDNode *N, SelectionDAG &DAG) { 15963 const unsigned OffsetPos = 4; 15964 SDValue Offset = N->getOperand(OffsetPos); 15965 15966 // Not an unpacked vector, bail out. 15967 if (Offset.getValueType().getSimpleVT().SimpleTy != MVT::nxv2i32) 15968 return SDValue(); 15969 15970 // Extend the unpacked offset vector to 64-bit lanes. 15971 SDLoc DL(N); 15972 Offset = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::nxv2i64, Offset); 15973 SmallVector<SDValue, 5> Ops(N->op_begin(), N->op_end()); 15974 // Replace the offset operand with the 64-bit one. 15975 Ops[OffsetPos] = Offset; 15976 15977 return DAG.getNode(N->getOpcode(), DL, DAG.getVTList(MVT::Other), Ops); 15978 } 15979 15980 /// Combines a node carrying the intrinsic 15981 /// `aarch64_sve_prf<T>_gather_scalar_offset` into a node that uses 15982 /// `aarch64_sve_prfb_gather_uxtw_index` when the scalar offset passed to 15983 /// `aarch64_sve_prf<T>_gather_scalar_offset` is not a valid immediate for the 15984 /// sve gather prefetch instruction with vector plus immediate addressing mode. 15985 static SDValue combineSVEPrefetchVecBaseImmOff(SDNode *N, SelectionDAG &DAG, 15986 unsigned ScalarSizeInBytes) { 15987 const unsigned ImmPos = 4, OffsetPos = 3; 15988 // No need to combine the node if the immediate is valid... 15989 if (isValidImmForSVEVecImmAddrMode(N->getOperand(ImmPos), ScalarSizeInBytes)) 15990 return SDValue(); 15991 15992 // ...otherwise swap the offset base with the offset... 15993 SmallVector<SDValue, 5> Ops(N->op_begin(), N->op_end()); 15994 std::swap(Ops[ImmPos], Ops[OffsetPos]); 15995 // ...and remap the intrinsic `aarch64_sve_prf<T>_gather_scalar_offset` to 15996 // `aarch64_sve_prfb_gather_uxtw_index`. 15997 SDLoc DL(N); 15998 Ops[1] = DAG.getConstant(Intrinsic::aarch64_sve_prfb_gather_uxtw_index, DL, 15999 MVT::i64); 16000 16001 return DAG.getNode(N->getOpcode(), DL, DAG.getVTList(MVT::Other), Ops); 16002 } 16003 16004 // Return true if the vector operation can guarantee only the first lane of its 16005 // result contains data, with all bits in other lanes set to zero. 16006 static bool isLanes1toNKnownZero(SDValue Op) { 16007 switch (Op.getOpcode()) { 16008 default: 16009 return false; 16010 case AArch64ISD::ANDV_PRED: 16011 case AArch64ISD::EORV_PRED: 16012 case AArch64ISD::FADDA_PRED: 16013 case AArch64ISD::FADDV_PRED: 16014 case AArch64ISD::FMAXNMV_PRED: 16015 case AArch64ISD::FMAXV_PRED: 16016 case AArch64ISD::FMINNMV_PRED: 16017 case AArch64ISD::FMINV_PRED: 16018 case AArch64ISD::ORV_PRED: 16019 case AArch64ISD::SADDV_PRED: 16020 case AArch64ISD::SMAXV_PRED: 16021 case AArch64ISD::SMINV_PRED: 16022 case AArch64ISD::UADDV_PRED: 16023 case AArch64ISD::UMAXV_PRED: 16024 case AArch64ISD::UMINV_PRED: 16025 return true; 16026 } 16027 } 16028 16029 static SDValue removeRedundantInsertVectorElt(SDNode *N) { 16030 assert(N->getOpcode() == ISD::INSERT_VECTOR_ELT && "Unexpected node!"); 16031 SDValue InsertVec = N->getOperand(0); 16032 SDValue InsertElt = N->getOperand(1); 16033 SDValue InsertIdx = N->getOperand(2); 16034 16035 // We only care about inserts into the first element... 16036 if (!isNullConstant(InsertIdx)) 16037 return SDValue(); 16038 // ...of a zero'd vector... 16039 if (!ISD::isConstantSplatVectorAllZeros(InsertVec.getNode())) 16040 return SDValue(); 16041 // ...where the inserted data was previously extracted... 16042 if (InsertElt.getOpcode() != ISD::EXTRACT_VECTOR_ELT) 16043 return SDValue(); 16044 16045 SDValue ExtractVec = InsertElt.getOperand(0); 16046 SDValue ExtractIdx = InsertElt.getOperand(1); 16047 16048 // ...from the first element of a vector. 16049 if (!isNullConstant(ExtractIdx)) 16050 return SDValue(); 16051 16052 // If we get here we are effectively trying to zero lanes 1-N of a vector. 16053 16054 // Ensure there's no type conversion going on. 16055 if (N->getValueType(0) != ExtractVec.getValueType()) 16056 return SDValue(); 16057 16058 if (!isLanes1toNKnownZero(ExtractVec)) 16059 return SDValue(); 16060 16061 // The explicit zeroing is redundant. 16062 return ExtractVec; 16063 } 16064 16065 static SDValue 16066 performInsertVectorEltCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) { 16067 if (SDValue Res = removeRedundantInsertVectorElt(N)) 16068 return Res; 16069 16070 return performPostLD1Combine(N, DCI, true); 16071 } 16072 16073 SDValue AArch64TargetLowering::PerformDAGCombine(SDNode *N, 16074 DAGCombinerInfo &DCI) const { 16075 SelectionDAG &DAG = DCI.DAG; 16076 switch (N->getOpcode()) { 16077 default: 16078 LLVM_DEBUG(dbgs() << "Custom combining: skipping\n"); 16079 break; 16080 case ISD::ABS: 16081 return performABSCombine(N, DAG, DCI, Subtarget); 16082 case ISD::ADD: 16083 case ISD::SUB: 16084 return performAddSubCombine(N, DCI, DAG); 16085 case ISD::XOR: 16086 return performXorCombine(N, DAG, DCI, Subtarget); 16087 case ISD::MUL: 16088 return performMulCombine(N, DAG, DCI, Subtarget); 16089 case ISD::SINT_TO_FP: 16090 case ISD::UINT_TO_FP: 16091 return performIntToFpCombine(N, DAG, Subtarget); 16092 case ISD::FP_TO_SINT: 16093 case ISD::FP_TO_UINT: 16094 return performFpToIntCombine(N, DAG, DCI, Subtarget); 16095 case ISD::FDIV: 16096 return performFDivCombine(N, DAG, DCI, Subtarget); 16097 case ISD::OR: 16098 return performORCombine(N, DCI, Subtarget); 16099 case ISD::AND: 16100 return performANDCombine(N, DCI); 16101 case ISD::SRL: 16102 return performSRLCombine(N, DCI); 16103 case ISD::INTRINSIC_WO_CHAIN: 16104 return performIntrinsicCombine(N, DCI, Subtarget); 16105 case ISD::ANY_EXTEND: 16106 case ISD::ZERO_EXTEND: 16107 case ISD::SIGN_EXTEND: 16108 return performExtendCombine(N, DCI, DAG); 16109 case ISD::SIGN_EXTEND_INREG: 16110 return performSignExtendInRegCombine(N, DCI, DAG); 16111 case ISD::TRUNCATE: 16112 return performVectorTruncateCombine(N, DCI, DAG); 16113 case ISD::CONCAT_VECTORS: 16114 return performConcatVectorsCombine(N, DCI, DAG); 16115 case ISD::SELECT: 16116 return performSelectCombine(N, DCI); 16117 case ISD::VSELECT: 16118 return performVSelectCombine(N, DCI.DAG); 16119 case ISD::LOAD: 16120 if (performTBISimplification(N->getOperand(1), DCI, DAG)) 16121 return SDValue(N, 0); 16122 break; 16123 case ISD::STORE: 16124 return performSTORECombine(N, DCI, DAG, Subtarget); 16125 case AArch64ISD::BRCOND: 16126 return performBRCONDCombine(N, DCI, DAG); 16127 case AArch64ISD::TBNZ: 16128 case AArch64ISD::TBZ: 16129 return performTBZCombine(N, DCI, DAG); 16130 case AArch64ISD::CSEL: 16131 return performCSELCombine(N, DCI, DAG); 16132 case AArch64ISD::DUP: 16133 return performPostLD1Combine(N, DCI, false); 16134 case AArch64ISD::NVCAST: 16135 return performNVCASTCombine(N); 16136 case AArch64ISD::UZP1: 16137 return performUzpCombine(N, DAG); 16138 case AArch64ISD::GLD1_MERGE_ZERO: 16139 case AArch64ISD::GLD1_SCALED_MERGE_ZERO: 16140 case AArch64ISD::GLD1_UXTW_MERGE_ZERO: 16141 case AArch64ISD::GLD1_SXTW_MERGE_ZERO: 16142 case AArch64ISD::GLD1_UXTW_SCALED_MERGE_ZERO: 16143 case AArch64ISD::GLD1_SXTW_SCALED_MERGE_ZERO: 16144 case AArch64ISD::GLD1_IMM_MERGE_ZERO: 16145 case AArch64ISD::GLD1S_MERGE_ZERO: 16146 case AArch64ISD::GLD1S_SCALED_MERGE_ZERO: 16147 case AArch64ISD::GLD1S_UXTW_MERGE_ZERO: 16148 case AArch64ISD::GLD1S_SXTW_MERGE_ZERO: 16149 case AArch64ISD::GLD1S_UXTW_SCALED_MERGE_ZERO: 16150 case AArch64ISD::GLD1S_SXTW_SCALED_MERGE_ZERO: 16151 case AArch64ISD::GLD1S_IMM_MERGE_ZERO: 16152 return performGLD1Combine(N, DAG); 16153 case AArch64ISD::VASHR: 16154 case AArch64ISD::VLSHR: 16155 return performVectorShiftCombine(N, *this, DCI); 16156 case ISD::INSERT_VECTOR_ELT: 16157 return performInsertVectorEltCombine(N, DCI); 16158 case ISD::EXTRACT_VECTOR_ELT: 16159 return performExtractVectorEltCombine(N, DAG); 16160 case ISD::VECREDUCE_ADD: 16161 return performVecReduceAddCombine(N, DCI.DAG, Subtarget); 16162 case ISD::INTRINSIC_VOID: 16163 case ISD::INTRINSIC_W_CHAIN: 16164 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) { 16165 case Intrinsic::aarch64_sve_prfb_gather_scalar_offset: 16166 return combineSVEPrefetchVecBaseImmOff(N, DAG, 1 /*=ScalarSizeInBytes*/); 16167 case Intrinsic::aarch64_sve_prfh_gather_scalar_offset: 16168 return combineSVEPrefetchVecBaseImmOff(N, DAG, 2 /*=ScalarSizeInBytes*/); 16169 case Intrinsic::aarch64_sve_prfw_gather_scalar_offset: 16170 return combineSVEPrefetchVecBaseImmOff(N, DAG, 4 /*=ScalarSizeInBytes*/); 16171 case Intrinsic::aarch64_sve_prfd_gather_scalar_offset: 16172 return combineSVEPrefetchVecBaseImmOff(N, DAG, 8 /*=ScalarSizeInBytes*/); 16173 case Intrinsic::aarch64_sve_prfb_gather_uxtw_index: 16174 case Intrinsic::aarch64_sve_prfb_gather_sxtw_index: 16175 case Intrinsic::aarch64_sve_prfh_gather_uxtw_index: 16176 case Intrinsic::aarch64_sve_prfh_gather_sxtw_index: 16177 case Intrinsic::aarch64_sve_prfw_gather_uxtw_index: 16178 case Intrinsic::aarch64_sve_prfw_gather_sxtw_index: 16179 case Intrinsic::aarch64_sve_prfd_gather_uxtw_index: 16180 case Intrinsic::aarch64_sve_prfd_gather_sxtw_index: 16181 return legalizeSVEGatherPrefetchOffsVec(N, DAG); 16182 case Intrinsic::aarch64_neon_ld2: 16183 case Intrinsic::aarch64_neon_ld3: 16184 case Intrinsic::aarch64_neon_ld4: 16185 case Intrinsic::aarch64_neon_ld1x2: 16186 case Intrinsic::aarch64_neon_ld1x3: 16187 case Intrinsic::aarch64_neon_ld1x4: 16188 case Intrinsic::aarch64_neon_ld2lane: 16189 case Intrinsic::aarch64_neon_ld3lane: 16190 case Intrinsic::aarch64_neon_ld4lane: 16191 case Intrinsic::aarch64_neon_ld2r: 16192 case Intrinsic::aarch64_neon_ld3r: 16193 case Intrinsic::aarch64_neon_ld4r: 16194 case Intrinsic::aarch64_neon_st2: 16195 case Intrinsic::aarch64_neon_st3: 16196 case Intrinsic::aarch64_neon_st4: 16197 case Intrinsic::aarch64_neon_st1x2: 16198 case Intrinsic::aarch64_neon_st1x3: 16199 case Intrinsic::aarch64_neon_st1x4: 16200 case Intrinsic::aarch64_neon_st2lane: 16201 case Intrinsic::aarch64_neon_st3lane: 16202 case Intrinsic::aarch64_neon_st4lane: 16203 return performNEONPostLDSTCombine(N, DCI, DAG); 16204 case Intrinsic::aarch64_sve_ldnt1: 16205 return performLDNT1Combine(N, DAG); 16206 case Intrinsic::aarch64_sve_ld1rq: 16207 return performLD1ReplicateCombine<AArch64ISD::LD1RQ_MERGE_ZERO>(N, DAG); 16208 case Intrinsic::aarch64_sve_ld1ro: 16209 return performLD1ReplicateCombine<AArch64ISD::LD1RO_MERGE_ZERO>(N, DAG); 16210 case Intrinsic::aarch64_sve_ldnt1_gather_scalar_offset: 16211 return performGatherLoadCombine(N, DAG, AArch64ISD::GLDNT1_MERGE_ZERO); 16212 case Intrinsic::aarch64_sve_ldnt1_gather: 16213 return performGatherLoadCombine(N, DAG, AArch64ISD::GLDNT1_MERGE_ZERO); 16214 case Intrinsic::aarch64_sve_ldnt1_gather_index: 16215 return performGatherLoadCombine(N, DAG, 16216 AArch64ISD::GLDNT1_INDEX_MERGE_ZERO); 16217 case Intrinsic::aarch64_sve_ldnt1_gather_uxtw: 16218 return performGatherLoadCombine(N, DAG, AArch64ISD::GLDNT1_MERGE_ZERO); 16219 case Intrinsic::aarch64_sve_ld1: 16220 return performLD1Combine(N, DAG, AArch64ISD::LD1_MERGE_ZERO); 16221 case Intrinsic::aarch64_sve_ldnf1: 16222 return performLD1Combine(N, DAG, AArch64ISD::LDNF1_MERGE_ZERO); 16223 case Intrinsic::aarch64_sve_ldff1: 16224 return performLD1Combine(N, DAG, AArch64ISD::LDFF1_MERGE_ZERO); 16225 case Intrinsic::aarch64_sve_st1: 16226 return performST1Combine(N, DAG); 16227 case Intrinsic::aarch64_sve_stnt1: 16228 return performSTNT1Combine(N, DAG); 16229 case Intrinsic::aarch64_sve_stnt1_scatter_scalar_offset: 16230 return performScatterStoreCombine(N, DAG, AArch64ISD::SSTNT1_PRED); 16231 case Intrinsic::aarch64_sve_stnt1_scatter_uxtw: 16232 return performScatterStoreCombine(N, DAG, AArch64ISD::SSTNT1_PRED); 16233 case Intrinsic::aarch64_sve_stnt1_scatter: 16234 return performScatterStoreCombine(N, DAG, AArch64ISD::SSTNT1_PRED); 16235 case Intrinsic::aarch64_sve_stnt1_scatter_index: 16236 return performScatterStoreCombine(N, DAG, AArch64ISD::SSTNT1_INDEX_PRED); 16237 case Intrinsic::aarch64_sve_ld1_gather: 16238 return performGatherLoadCombine(N, DAG, AArch64ISD::GLD1_MERGE_ZERO); 16239 case Intrinsic::aarch64_sve_ld1_gather_index: 16240 return performGatherLoadCombine(N, DAG, 16241 AArch64ISD::GLD1_SCALED_MERGE_ZERO); 16242 case Intrinsic::aarch64_sve_ld1_gather_sxtw: 16243 return performGatherLoadCombine(N, DAG, AArch64ISD::GLD1_SXTW_MERGE_ZERO, 16244 /*OnlyPackedOffsets=*/false); 16245 case Intrinsic::aarch64_sve_ld1_gather_uxtw: 16246 return performGatherLoadCombine(N, DAG, AArch64ISD::GLD1_UXTW_MERGE_ZERO, 16247 /*OnlyPackedOffsets=*/false); 16248 case Intrinsic::aarch64_sve_ld1_gather_sxtw_index: 16249 return performGatherLoadCombine(N, DAG, 16250 AArch64ISD::GLD1_SXTW_SCALED_MERGE_ZERO, 16251 /*OnlyPackedOffsets=*/false); 16252 case Intrinsic::aarch64_sve_ld1_gather_uxtw_index: 16253 return performGatherLoadCombine(N, DAG, 16254 AArch64ISD::GLD1_UXTW_SCALED_MERGE_ZERO, 16255 /*OnlyPackedOffsets=*/false); 16256 case Intrinsic::aarch64_sve_ld1_gather_scalar_offset: 16257 return performGatherLoadCombine(N, DAG, AArch64ISD::GLD1_IMM_MERGE_ZERO); 16258 case Intrinsic::aarch64_sve_ldff1_gather: 16259 return performGatherLoadCombine(N, DAG, AArch64ISD::GLDFF1_MERGE_ZERO); 16260 case Intrinsic::aarch64_sve_ldff1_gather_index: 16261 return performGatherLoadCombine(N, DAG, 16262 AArch64ISD::GLDFF1_SCALED_MERGE_ZERO); 16263 case Intrinsic::aarch64_sve_ldff1_gather_sxtw: 16264 return performGatherLoadCombine(N, DAG, 16265 AArch64ISD::GLDFF1_SXTW_MERGE_ZERO, 16266 /*OnlyPackedOffsets=*/false); 16267 case Intrinsic::aarch64_sve_ldff1_gather_uxtw: 16268 return performGatherLoadCombine(N, DAG, 16269 AArch64ISD::GLDFF1_UXTW_MERGE_ZERO, 16270 /*OnlyPackedOffsets=*/false); 16271 case Intrinsic::aarch64_sve_ldff1_gather_sxtw_index: 16272 return performGatherLoadCombine(N, DAG, 16273 AArch64ISD::GLDFF1_SXTW_SCALED_MERGE_ZERO, 16274 /*OnlyPackedOffsets=*/false); 16275 case Intrinsic::aarch64_sve_ldff1_gather_uxtw_index: 16276 return performGatherLoadCombine(N, DAG, 16277 AArch64ISD::GLDFF1_UXTW_SCALED_MERGE_ZERO, 16278 /*OnlyPackedOffsets=*/false); 16279 case Intrinsic::aarch64_sve_ldff1_gather_scalar_offset: 16280 return performGatherLoadCombine(N, DAG, 16281 AArch64ISD::GLDFF1_IMM_MERGE_ZERO); 16282 case Intrinsic::aarch64_sve_st1_scatter: 16283 return performScatterStoreCombine(N, DAG, AArch64ISD::SST1_PRED); 16284 case Intrinsic::aarch64_sve_st1_scatter_index: 16285 return performScatterStoreCombine(N, DAG, AArch64ISD::SST1_SCALED_PRED); 16286 case Intrinsic::aarch64_sve_st1_scatter_sxtw: 16287 return performScatterStoreCombine(N, DAG, AArch64ISD::SST1_SXTW_PRED, 16288 /*OnlyPackedOffsets=*/false); 16289 case Intrinsic::aarch64_sve_st1_scatter_uxtw: 16290 return performScatterStoreCombine(N, DAG, AArch64ISD::SST1_UXTW_PRED, 16291 /*OnlyPackedOffsets=*/false); 16292 case Intrinsic::aarch64_sve_st1_scatter_sxtw_index: 16293 return performScatterStoreCombine(N, DAG, 16294 AArch64ISD::SST1_SXTW_SCALED_PRED, 16295 /*OnlyPackedOffsets=*/false); 16296 case Intrinsic::aarch64_sve_st1_scatter_uxtw_index: 16297 return performScatterStoreCombine(N, DAG, 16298 AArch64ISD::SST1_UXTW_SCALED_PRED, 16299 /*OnlyPackedOffsets=*/false); 16300 case Intrinsic::aarch64_sve_st1_scatter_scalar_offset: 16301 return performScatterStoreCombine(N, DAG, AArch64ISD::SST1_IMM_PRED); 16302 case Intrinsic::aarch64_sve_tuple_get: { 16303 SDLoc DL(N); 16304 SDValue Chain = N->getOperand(0); 16305 SDValue Src1 = N->getOperand(2); 16306 SDValue Idx = N->getOperand(3); 16307 16308 uint64_t IdxConst = cast<ConstantSDNode>(Idx)->getZExtValue(); 16309 EVT ResVT = N->getValueType(0); 16310 uint64_t NumLanes = ResVT.getVectorElementCount().getKnownMinValue(); 16311 SDValue ExtIdx = DAG.getVectorIdxConstant(IdxConst * NumLanes, DL); 16312 SDValue Val = 16313 DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ResVT, Src1, ExtIdx); 16314 return DAG.getMergeValues({Val, Chain}, DL); 16315 } 16316 case Intrinsic::aarch64_sve_tuple_set: { 16317 SDLoc DL(N); 16318 SDValue Chain = N->getOperand(0); 16319 SDValue Tuple = N->getOperand(2); 16320 SDValue Idx = N->getOperand(3); 16321 SDValue Vec = N->getOperand(4); 16322 16323 EVT TupleVT = Tuple.getValueType(); 16324 uint64_t TupleLanes = TupleVT.getVectorElementCount().getKnownMinValue(); 16325 16326 uint64_t IdxConst = cast<ConstantSDNode>(Idx)->getZExtValue(); 16327 uint64_t NumLanes = 16328 Vec.getValueType().getVectorElementCount().getKnownMinValue(); 16329 16330 if ((TupleLanes % NumLanes) != 0) 16331 report_fatal_error("invalid tuple vector!"); 16332 16333 uint64_t NumVecs = TupleLanes / NumLanes; 16334 16335 SmallVector<SDValue, 4> Opnds; 16336 for (unsigned I = 0; I < NumVecs; ++I) { 16337 if (I == IdxConst) 16338 Opnds.push_back(Vec); 16339 else { 16340 SDValue ExtIdx = DAG.getVectorIdxConstant(I * NumLanes, DL); 16341 Opnds.push_back(DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, 16342 Vec.getValueType(), Tuple, ExtIdx)); 16343 } 16344 } 16345 SDValue Concat = 16346 DAG.getNode(ISD::CONCAT_VECTORS, DL, Tuple.getValueType(), Opnds); 16347 return DAG.getMergeValues({Concat, Chain}, DL); 16348 } 16349 case Intrinsic::aarch64_sve_tuple_create2: 16350 case Intrinsic::aarch64_sve_tuple_create3: 16351 case Intrinsic::aarch64_sve_tuple_create4: { 16352 SDLoc DL(N); 16353 SDValue Chain = N->getOperand(0); 16354 16355 SmallVector<SDValue, 4> Opnds; 16356 for (unsigned I = 2; I < N->getNumOperands(); ++I) 16357 Opnds.push_back(N->getOperand(I)); 16358 16359 EVT VT = Opnds[0].getValueType(); 16360 EVT EltVT = VT.getVectorElementType(); 16361 EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT, 16362 VT.getVectorElementCount() * 16363 (N->getNumOperands() - 2)); 16364 SDValue Concat = DAG.getNode(ISD::CONCAT_VECTORS, DL, DestVT, Opnds); 16365 return DAG.getMergeValues({Concat, Chain}, DL); 16366 } 16367 case Intrinsic::aarch64_sve_ld2: 16368 case Intrinsic::aarch64_sve_ld3: 16369 case Intrinsic::aarch64_sve_ld4: { 16370 SDLoc DL(N); 16371 SDValue Chain = N->getOperand(0); 16372 SDValue Mask = N->getOperand(2); 16373 SDValue BasePtr = N->getOperand(3); 16374 SDValue LoadOps[] = {Chain, Mask, BasePtr}; 16375 unsigned IntrinsicID = 16376 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue(); 16377 SDValue Result = 16378 LowerSVEStructLoad(IntrinsicID, LoadOps, N->getValueType(0), DAG, DL); 16379 return DAG.getMergeValues({Result, Chain}, DL); 16380 } 16381 case Intrinsic::aarch64_rndr: 16382 case Intrinsic::aarch64_rndrrs: { 16383 unsigned IntrinsicID = 16384 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue(); 16385 auto Register = 16386 (IntrinsicID == Intrinsic::aarch64_rndr ? AArch64SysReg::RNDR 16387 : AArch64SysReg::RNDRRS); 16388 SDLoc DL(N); 16389 SDValue A = DAG.getNode( 16390 AArch64ISD::MRS, DL, DAG.getVTList(MVT::i64, MVT::Glue, MVT::Other), 16391 N->getOperand(0), DAG.getConstant(Register, DL, MVT::i64)); 16392 SDValue B = DAG.getNode( 16393 AArch64ISD::CSINC, DL, MVT::i32, DAG.getConstant(0, DL, MVT::i32), 16394 DAG.getConstant(0, DL, MVT::i32), 16395 DAG.getConstant(AArch64CC::NE, DL, MVT::i32), A.getValue(1)); 16396 return DAG.getMergeValues( 16397 {A, DAG.getZExtOrTrunc(B, DL, MVT::i1), A.getValue(2)}, DL); 16398 } 16399 default: 16400 break; 16401 } 16402 break; 16403 case ISD::GlobalAddress: 16404 return performGlobalAddressCombine(N, DAG, Subtarget, getTargetMachine()); 16405 } 16406 return SDValue(); 16407 } 16408 16409 // Check if the return value is used as only a return value, as otherwise 16410 // we can't perform a tail-call. In particular, we need to check for 16411 // target ISD nodes that are returns and any other "odd" constructs 16412 // that the generic analysis code won't necessarily catch. 16413 bool AArch64TargetLowering::isUsedByReturnOnly(SDNode *N, 16414 SDValue &Chain) const { 16415 if (N->getNumValues() != 1) 16416 return false; 16417 if (!N->hasNUsesOfValue(1, 0)) 16418 return false; 16419 16420 SDValue TCChain = Chain; 16421 SDNode *Copy = *N->use_begin(); 16422 if (Copy->getOpcode() == ISD::CopyToReg) { 16423 // If the copy has a glue operand, we conservatively assume it isn't safe to 16424 // perform a tail call. 16425 if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() == 16426 MVT::Glue) 16427 return false; 16428 TCChain = Copy->getOperand(0); 16429 } else if (Copy->getOpcode() != ISD::FP_EXTEND) 16430 return false; 16431 16432 bool HasRet = false; 16433 for (SDNode *Node : Copy->uses()) { 16434 if (Node->getOpcode() != AArch64ISD::RET_FLAG) 16435 return false; 16436 HasRet = true; 16437 } 16438 16439 if (!HasRet) 16440 return false; 16441 16442 Chain = TCChain; 16443 return true; 16444 } 16445 16446 // Return whether the an instruction can potentially be optimized to a tail 16447 // call. This will cause the optimizers to attempt to move, or duplicate, 16448 // return instructions to help enable tail call optimizations for this 16449 // instruction. 16450 bool AArch64TargetLowering::mayBeEmittedAsTailCall(const CallInst *CI) const { 16451 return CI->isTailCall(); 16452 } 16453 16454 bool AArch64TargetLowering::getIndexedAddressParts(SDNode *Op, SDValue &Base, 16455 SDValue &Offset, 16456 ISD::MemIndexedMode &AM, 16457 bool &IsInc, 16458 SelectionDAG &DAG) const { 16459 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB) 16460 return false; 16461 16462 Base = Op->getOperand(0); 16463 // All of the indexed addressing mode instructions take a signed 16464 // 9 bit immediate offset. 16465 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) { 16466 int64_t RHSC = RHS->getSExtValue(); 16467 if (Op->getOpcode() == ISD::SUB) 16468 RHSC = -(uint64_t)RHSC; 16469 if (!isInt<9>(RHSC)) 16470 return false; 16471 IsInc = (Op->getOpcode() == ISD::ADD); 16472 Offset = Op->getOperand(1); 16473 return true; 16474 } 16475 return false; 16476 } 16477 16478 bool AArch64TargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base, 16479 SDValue &Offset, 16480 ISD::MemIndexedMode &AM, 16481 SelectionDAG &DAG) const { 16482 EVT VT; 16483 SDValue Ptr; 16484 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { 16485 VT = LD->getMemoryVT(); 16486 Ptr = LD->getBasePtr(); 16487 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { 16488 VT = ST->getMemoryVT(); 16489 Ptr = ST->getBasePtr(); 16490 } else 16491 return false; 16492 16493 bool IsInc; 16494 if (!getIndexedAddressParts(Ptr.getNode(), Base, Offset, AM, IsInc, DAG)) 16495 return false; 16496 AM = IsInc ? ISD::PRE_INC : ISD::PRE_DEC; 16497 return true; 16498 } 16499 16500 bool AArch64TargetLowering::getPostIndexedAddressParts( 16501 SDNode *N, SDNode *Op, SDValue &Base, SDValue &Offset, 16502 ISD::MemIndexedMode &AM, SelectionDAG &DAG) const { 16503 EVT VT; 16504 SDValue Ptr; 16505 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { 16506 VT = LD->getMemoryVT(); 16507 Ptr = LD->getBasePtr(); 16508 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { 16509 VT = ST->getMemoryVT(); 16510 Ptr = ST->getBasePtr(); 16511 } else 16512 return false; 16513 16514 bool IsInc; 16515 if (!getIndexedAddressParts(Op, Base, Offset, AM, IsInc, DAG)) 16516 return false; 16517 // Post-indexing updates the base, so it's not a valid transform 16518 // if that's not the same as the load's pointer. 16519 if (Ptr != Base) 16520 return false; 16521 AM = IsInc ? ISD::POST_INC : ISD::POST_DEC; 16522 return true; 16523 } 16524 16525 static void ReplaceBITCASTResults(SDNode *N, SmallVectorImpl<SDValue> &Results, 16526 SelectionDAG &DAG) { 16527 SDLoc DL(N); 16528 SDValue Op = N->getOperand(0); 16529 16530 if (N->getValueType(0) != MVT::i16 || 16531 (Op.getValueType() != MVT::f16 && Op.getValueType() != MVT::bf16)) 16532 return; 16533 16534 Op = SDValue( 16535 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, DL, MVT::f32, 16536 DAG.getUNDEF(MVT::i32), Op, 16537 DAG.getTargetConstant(AArch64::hsub, DL, MVT::i32)), 16538 0); 16539 Op = DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op); 16540 Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, Op)); 16541 } 16542 16543 static void ReplaceReductionResults(SDNode *N, 16544 SmallVectorImpl<SDValue> &Results, 16545 SelectionDAG &DAG, unsigned InterOp, 16546 unsigned AcrossOp) { 16547 EVT LoVT, HiVT; 16548 SDValue Lo, Hi; 16549 SDLoc dl(N); 16550 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0)); 16551 std::tie(Lo, Hi) = DAG.SplitVectorOperand(N, 0); 16552 SDValue InterVal = DAG.getNode(InterOp, dl, LoVT, Lo, Hi); 16553 SDValue SplitVal = DAG.getNode(AcrossOp, dl, LoVT, InterVal); 16554 Results.push_back(SplitVal); 16555 } 16556 16557 static std::pair<SDValue, SDValue> splitInt128(SDValue N, SelectionDAG &DAG) { 16558 SDLoc DL(N); 16559 SDValue Lo = DAG.getNode(ISD::TRUNCATE, DL, MVT::i64, N); 16560 SDValue Hi = DAG.getNode(ISD::TRUNCATE, DL, MVT::i64, 16561 DAG.getNode(ISD::SRL, DL, MVT::i128, N, 16562 DAG.getConstant(64, DL, MVT::i64))); 16563 return std::make_pair(Lo, Hi); 16564 } 16565 16566 void AArch64TargetLowering::ReplaceExtractSubVectorResults( 16567 SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const { 16568 SDValue In = N->getOperand(0); 16569 EVT InVT = In.getValueType(); 16570 16571 // Common code will handle these just fine. 16572 if (!InVT.isScalableVector() || !InVT.isInteger()) 16573 return; 16574 16575 SDLoc DL(N); 16576 EVT VT = N->getValueType(0); 16577 16578 // The following checks bail if this is not a halving operation. 16579 16580 ElementCount ResEC = VT.getVectorElementCount(); 16581 16582 if (InVT.getVectorElementCount() != (ResEC * 2)) 16583 return; 16584 16585 auto *CIndex = dyn_cast<ConstantSDNode>(N->getOperand(1)); 16586 if (!CIndex) 16587 return; 16588 16589 unsigned Index = CIndex->getZExtValue(); 16590 if ((Index != 0) && (Index != ResEC.getKnownMinValue())) 16591 return; 16592 16593 unsigned Opcode = (Index == 0) ? AArch64ISD::UUNPKLO : AArch64ISD::UUNPKHI; 16594 EVT ExtendedHalfVT = VT.widenIntegerVectorElementType(*DAG.getContext()); 16595 16596 SDValue Half = DAG.getNode(Opcode, DL, ExtendedHalfVT, N->getOperand(0)); 16597 Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, Half)); 16598 } 16599 16600 // Create an even/odd pair of X registers holding integer value V. 16601 static SDValue createGPRPairNode(SelectionDAG &DAG, SDValue V) { 16602 SDLoc dl(V.getNode()); 16603 SDValue VLo = DAG.getAnyExtOrTrunc(V, dl, MVT::i64); 16604 SDValue VHi = DAG.getAnyExtOrTrunc( 16605 DAG.getNode(ISD::SRL, dl, MVT::i128, V, DAG.getConstant(64, dl, MVT::i64)), 16606 dl, MVT::i64); 16607 if (DAG.getDataLayout().isBigEndian()) 16608 std::swap (VLo, VHi); 16609 SDValue RegClass = 16610 DAG.getTargetConstant(AArch64::XSeqPairsClassRegClassID, dl, MVT::i32); 16611 SDValue SubReg0 = DAG.getTargetConstant(AArch64::sube64, dl, MVT::i32); 16612 SDValue SubReg1 = DAG.getTargetConstant(AArch64::subo64, dl, MVT::i32); 16613 const SDValue Ops[] = { RegClass, VLo, SubReg0, VHi, SubReg1 }; 16614 return SDValue( 16615 DAG.getMachineNode(TargetOpcode::REG_SEQUENCE, dl, MVT::Untyped, Ops), 0); 16616 } 16617 16618 static void ReplaceCMP_SWAP_128Results(SDNode *N, 16619 SmallVectorImpl<SDValue> &Results, 16620 SelectionDAG &DAG, 16621 const AArch64Subtarget *Subtarget) { 16622 assert(N->getValueType(0) == MVT::i128 && 16623 "AtomicCmpSwap on types less than 128 should be legal"); 16624 16625 if (Subtarget->hasLSE() || Subtarget->outlineAtomics()) { 16626 // LSE has a 128-bit compare and swap (CASP), but i128 is not a legal type, 16627 // so lower it here, wrapped in REG_SEQUENCE and EXTRACT_SUBREG. 16628 SDValue Ops[] = { 16629 createGPRPairNode(DAG, N->getOperand(2)), // Compare value 16630 createGPRPairNode(DAG, N->getOperand(3)), // Store value 16631 N->getOperand(1), // Ptr 16632 N->getOperand(0), // Chain in 16633 }; 16634 16635 MachineMemOperand *MemOp = cast<MemSDNode>(N)->getMemOperand(); 16636 16637 unsigned Opcode; 16638 switch (MemOp->getOrdering()) { 16639 case AtomicOrdering::Monotonic: 16640 Opcode = AArch64::CASPX; 16641 break; 16642 case AtomicOrdering::Acquire: 16643 Opcode = AArch64::CASPAX; 16644 break; 16645 case AtomicOrdering::Release: 16646 Opcode = AArch64::CASPLX; 16647 break; 16648 case AtomicOrdering::AcquireRelease: 16649 case AtomicOrdering::SequentiallyConsistent: 16650 Opcode = AArch64::CASPALX; 16651 break; 16652 default: 16653 llvm_unreachable("Unexpected ordering!"); 16654 } 16655 16656 MachineSDNode *CmpSwap = DAG.getMachineNode( 16657 Opcode, SDLoc(N), DAG.getVTList(MVT::Untyped, MVT::Other), Ops); 16658 DAG.setNodeMemRefs(CmpSwap, {MemOp}); 16659 16660 unsigned SubReg1 = AArch64::sube64, SubReg2 = AArch64::subo64; 16661 if (DAG.getDataLayout().isBigEndian()) 16662 std::swap(SubReg1, SubReg2); 16663 SDValue Lo = DAG.getTargetExtractSubreg(SubReg1, SDLoc(N), MVT::i64, 16664 SDValue(CmpSwap, 0)); 16665 SDValue Hi = DAG.getTargetExtractSubreg(SubReg2, SDLoc(N), MVT::i64, 16666 SDValue(CmpSwap, 0)); 16667 Results.push_back( 16668 DAG.getNode(ISD::BUILD_PAIR, SDLoc(N), MVT::i128, Lo, Hi)); 16669 Results.push_back(SDValue(CmpSwap, 1)); // Chain out 16670 return; 16671 } 16672 16673 auto Desired = splitInt128(N->getOperand(2), DAG); 16674 auto New = splitInt128(N->getOperand(3), DAG); 16675 SDValue Ops[] = {N->getOperand(1), Desired.first, Desired.second, 16676 New.first, New.second, N->getOperand(0)}; 16677 SDNode *CmpSwap = DAG.getMachineNode( 16678 AArch64::CMP_SWAP_128, SDLoc(N), 16679 DAG.getVTList(MVT::i64, MVT::i64, MVT::i32, MVT::Other), Ops); 16680 16681 MachineMemOperand *MemOp = cast<MemSDNode>(N)->getMemOperand(); 16682 DAG.setNodeMemRefs(cast<MachineSDNode>(CmpSwap), {MemOp}); 16683 16684 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, SDLoc(N), MVT::i128, 16685 SDValue(CmpSwap, 0), SDValue(CmpSwap, 1))); 16686 Results.push_back(SDValue(CmpSwap, 3)); 16687 } 16688 16689 void AArch64TargetLowering::ReplaceNodeResults( 16690 SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const { 16691 switch (N->getOpcode()) { 16692 default: 16693 llvm_unreachable("Don't know how to custom expand this"); 16694 case ISD::BITCAST: 16695 ReplaceBITCASTResults(N, Results, DAG); 16696 return; 16697 case ISD::VECREDUCE_ADD: 16698 case ISD::VECREDUCE_SMAX: 16699 case ISD::VECREDUCE_SMIN: 16700 case ISD::VECREDUCE_UMAX: 16701 case ISD::VECREDUCE_UMIN: 16702 Results.push_back(LowerVECREDUCE(SDValue(N, 0), DAG)); 16703 return; 16704 16705 case ISD::CTPOP: 16706 if (SDValue Result = LowerCTPOP(SDValue(N, 0), DAG)) 16707 Results.push_back(Result); 16708 return; 16709 case AArch64ISD::SADDV: 16710 ReplaceReductionResults(N, Results, DAG, ISD::ADD, AArch64ISD::SADDV); 16711 return; 16712 case AArch64ISD::UADDV: 16713 ReplaceReductionResults(N, Results, DAG, ISD::ADD, AArch64ISD::UADDV); 16714 return; 16715 case AArch64ISD::SMINV: 16716 ReplaceReductionResults(N, Results, DAG, ISD::SMIN, AArch64ISD::SMINV); 16717 return; 16718 case AArch64ISD::UMINV: 16719 ReplaceReductionResults(N, Results, DAG, ISD::UMIN, AArch64ISD::UMINV); 16720 return; 16721 case AArch64ISD::SMAXV: 16722 ReplaceReductionResults(N, Results, DAG, ISD::SMAX, AArch64ISD::SMAXV); 16723 return; 16724 case AArch64ISD::UMAXV: 16725 ReplaceReductionResults(N, Results, DAG, ISD::UMAX, AArch64ISD::UMAXV); 16726 return; 16727 case ISD::FP_TO_UINT: 16728 case ISD::FP_TO_SINT: 16729 assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion"); 16730 // Let normal code take care of it by not adding anything to Results. 16731 return; 16732 case ISD::ATOMIC_CMP_SWAP: 16733 ReplaceCMP_SWAP_128Results(N, Results, DAG, Subtarget); 16734 return; 16735 case ISD::LOAD: { 16736 assert(SDValue(N, 0).getValueType() == MVT::i128 && 16737 "unexpected load's value type"); 16738 LoadSDNode *LoadNode = cast<LoadSDNode>(N); 16739 if (!LoadNode->isVolatile() || LoadNode->getMemoryVT() != MVT::i128) { 16740 // Non-volatile loads are optimized later in AArch64's load/store 16741 // optimizer. 16742 return; 16743 } 16744 16745 SDValue Result = DAG.getMemIntrinsicNode( 16746 AArch64ISD::LDP, SDLoc(N), 16747 DAG.getVTList({MVT::i64, MVT::i64, MVT::Other}), 16748 {LoadNode->getChain(), LoadNode->getBasePtr()}, LoadNode->getMemoryVT(), 16749 LoadNode->getMemOperand()); 16750 16751 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, SDLoc(N), MVT::i128, 16752 Result.getValue(0), Result.getValue(1)); 16753 Results.append({Pair, Result.getValue(2) /* Chain */}); 16754 return; 16755 } 16756 case ISD::EXTRACT_SUBVECTOR: 16757 ReplaceExtractSubVectorResults(N, Results, DAG); 16758 return; 16759 case ISD::INTRINSIC_WO_CHAIN: { 16760 EVT VT = N->getValueType(0); 16761 assert((VT == MVT::i8 || VT == MVT::i16) && 16762 "custom lowering for unexpected type"); 16763 16764 ConstantSDNode *CN = cast<ConstantSDNode>(N->getOperand(0)); 16765 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue()); 16766 switch (IntID) { 16767 default: 16768 return; 16769 case Intrinsic::aarch64_sve_clasta_n: { 16770 SDLoc DL(N); 16771 auto Op2 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, N->getOperand(2)); 16772 auto V = DAG.getNode(AArch64ISD::CLASTA_N, DL, MVT::i32, 16773 N->getOperand(1), Op2, N->getOperand(3)); 16774 Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, V)); 16775 return; 16776 } 16777 case Intrinsic::aarch64_sve_clastb_n: { 16778 SDLoc DL(N); 16779 auto Op2 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, N->getOperand(2)); 16780 auto V = DAG.getNode(AArch64ISD::CLASTB_N, DL, MVT::i32, 16781 N->getOperand(1), Op2, N->getOperand(3)); 16782 Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, V)); 16783 return; 16784 } 16785 case Intrinsic::aarch64_sve_lasta: { 16786 SDLoc DL(N); 16787 auto V = DAG.getNode(AArch64ISD::LASTA, DL, MVT::i32, 16788 N->getOperand(1), N->getOperand(2)); 16789 Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, V)); 16790 return; 16791 } 16792 case Intrinsic::aarch64_sve_lastb: { 16793 SDLoc DL(N); 16794 auto V = DAG.getNode(AArch64ISD::LASTB, DL, MVT::i32, 16795 N->getOperand(1), N->getOperand(2)); 16796 Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, V)); 16797 return; 16798 } 16799 } 16800 } 16801 } 16802 } 16803 16804 bool AArch64TargetLowering::useLoadStackGuardNode() const { 16805 if (Subtarget->isTargetAndroid() || Subtarget->isTargetFuchsia()) 16806 return TargetLowering::useLoadStackGuardNode(); 16807 return true; 16808 } 16809 16810 unsigned AArch64TargetLowering::combineRepeatedFPDivisors() const { 16811 // Combine multiple FDIVs with the same divisor into multiple FMULs by the 16812 // reciprocal if there are three or more FDIVs. 16813 return 3; 16814 } 16815 16816 TargetLoweringBase::LegalizeTypeAction 16817 AArch64TargetLowering::getPreferredVectorAction(MVT VT) const { 16818 // During type legalization, we prefer to widen v1i8, v1i16, v1i32 to v8i8, 16819 // v4i16, v2i32 instead of to promote. 16820 if (VT == MVT::v1i8 || VT == MVT::v1i16 || VT == MVT::v1i32 || 16821 VT == MVT::v1f32) 16822 return TypeWidenVector; 16823 16824 return TargetLoweringBase::getPreferredVectorAction(VT); 16825 } 16826 16827 // Loads and stores less than 128-bits are already atomic; ones above that 16828 // are doomed anyway, so defer to the default libcall and blame the OS when 16829 // things go wrong. 16830 bool AArch64TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const { 16831 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits(); 16832 return Size == 128; 16833 } 16834 16835 // Loads and stores less than 128-bits are already atomic; ones above that 16836 // are doomed anyway, so defer to the default libcall and blame the OS when 16837 // things go wrong. 16838 TargetLowering::AtomicExpansionKind 16839 AArch64TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const { 16840 unsigned Size = LI->getType()->getPrimitiveSizeInBits(); 16841 return Size == 128 ? AtomicExpansionKind::LLSC : AtomicExpansionKind::None; 16842 } 16843 16844 // For the real atomic operations, we have ldxr/stxr up to 128 bits, 16845 TargetLowering::AtomicExpansionKind 16846 AArch64TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const { 16847 if (AI->isFloatingPointOperation()) 16848 return AtomicExpansionKind::CmpXChg; 16849 16850 unsigned Size = AI->getType()->getPrimitiveSizeInBits(); 16851 if (Size > 128) return AtomicExpansionKind::None; 16852 16853 // Nand is not supported in LSE. 16854 // Leave 128 bits to LLSC or CmpXChg. 16855 if (AI->getOperation() != AtomicRMWInst::Nand && Size < 128) { 16856 if (Subtarget->hasLSE()) 16857 return AtomicExpansionKind::None; 16858 if (Subtarget->outlineAtomics()) { 16859 // [U]Min/[U]Max RWM atomics are used in __sync_fetch_ libcalls so far. 16860 // Don't outline them unless 16861 // (1) high level <atomic> support approved: 16862 // http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2020/p0493r1.pdf 16863 // (2) low level libgcc and compiler-rt support implemented by: 16864 // min/max outline atomics helpers 16865 if (AI->getOperation() != AtomicRMWInst::Min && 16866 AI->getOperation() != AtomicRMWInst::Max && 16867 AI->getOperation() != AtomicRMWInst::UMin && 16868 AI->getOperation() != AtomicRMWInst::UMax) { 16869 return AtomicExpansionKind::None; 16870 } 16871 } 16872 } 16873 16874 // At -O0, fast-regalloc cannot cope with the live vregs necessary to 16875 // implement atomicrmw without spilling. If the target address is also on the 16876 // stack and close enough to the spill slot, this can lead to a situation 16877 // where the monitor always gets cleared and the atomic operation can never 16878 // succeed. So at -O0 lower this operation to a CAS loop. 16879 if (getTargetMachine().getOptLevel() == CodeGenOpt::None) 16880 return AtomicExpansionKind::CmpXChg; 16881 16882 return AtomicExpansionKind::LLSC; 16883 } 16884 16885 TargetLowering::AtomicExpansionKind 16886 AArch64TargetLowering::shouldExpandAtomicCmpXchgInIR( 16887 AtomicCmpXchgInst *AI) const { 16888 // If subtarget has LSE, leave cmpxchg intact for codegen. 16889 if (Subtarget->hasLSE() || Subtarget->outlineAtomics()) 16890 return AtomicExpansionKind::None; 16891 // At -O0, fast-regalloc cannot cope with the live vregs necessary to 16892 // implement cmpxchg without spilling. If the address being exchanged is also 16893 // on the stack and close enough to the spill slot, this can lead to a 16894 // situation where the monitor always gets cleared and the atomic operation 16895 // can never succeed. So at -O0 we need a late-expanded pseudo-inst instead. 16896 if (getTargetMachine().getOptLevel() == CodeGenOpt::None) 16897 return AtomicExpansionKind::None; 16898 return AtomicExpansionKind::LLSC; 16899 } 16900 16901 Value *AArch64TargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr, 16902 AtomicOrdering Ord) const { 16903 Module *M = Builder.GetInsertBlock()->getParent()->getParent(); 16904 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType(); 16905 bool IsAcquire = isAcquireOrStronger(Ord); 16906 16907 // Since i128 isn't legal and intrinsics don't get type-lowered, the ldrexd 16908 // intrinsic must return {i64, i64} and we have to recombine them into a 16909 // single i128 here. 16910 if (ValTy->getPrimitiveSizeInBits() == 128) { 16911 Intrinsic::ID Int = 16912 IsAcquire ? Intrinsic::aarch64_ldaxp : Intrinsic::aarch64_ldxp; 16913 Function *Ldxr = Intrinsic::getDeclaration(M, Int); 16914 16915 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext())); 16916 Value *LoHi = Builder.CreateCall(Ldxr, Addr, "lohi"); 16917 16918 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo"); 16919 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi"); 16920 Lo = Builder.CreateZExt(Lo, ValTy, "lo64"); 16921 Hi = Builder.CreateZExt(Hi, ValTy, "hi64"); 16922 return Builder.CreateOr( 16923 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 64)), "val64"); 16924 } 16925 16926 Type *Tys[] = { Addr->getType() }; 16927 Intrinsic::ID Int = 16928 IsAcquire ? Intrinsic::aarch64_ldaxr : Intrinsic::aarch64_ldxr; 16929 Function *Ldxr = Intrinsic::getDeclaration(M, Int, Tys); 16930 16931 Type *EltTy = cast<PointerType>(Addr->getType())->getElementType(); 16932 16933 const DataLayout &DL = M->getDataLayout(); 16934 IntegerType *IntEltTy = Builder.getIntNTy(DL.getTypeSizeInBits(EltTy)); 16935 Value *Trunc = Builder.CreateTrunc(Builder.CreateCall(Ldxr, Addr), IntEltTy); 16936 16937 return Builder.CreateBitCast(Trunc, EltTy); 16938 } 16939 16940 void AArch64TargetLowering::emitAtomicCmpXchgNoStoreLLBalance( 16941 IRBuilder<> &Builder) const { 16942 Module *M = Builder.GetInsertBlock()->getParent()->getParent(); 16943 Builder.CreateCall(Intrinsic::getDeclaration(M, Intrinsic::aarch64_clrex)); 16944 } 16945 16946 Value *AArch64TargetLowering::emitStoreConditional(IRBuilder<> &Builder, 16947 Value *Val, Value *Addr, 16948 AtomicOrdering Ord) const { 16949 Module *M = Builder.GetInsertBlock()->getParent()->getParent(); 16950 bool IsRelease = isReleaseOrStronger(Ord); 16951 16952 // Since the intrinsics must have legal type, the i128 intrinsics take two 16953 // parameters: "i64, i64". We must marshal Val into the appropriate form 16954 // before the call. 16955 if (Val->getType()->getPrimitiveSizeInBits() == 128) { 16956 Intrinsic::ID Int = 16957 IsRelease ? Intrinsic::aarch64_stlxp : Intrinsic::aarch64_stxp; 16958 Function *Stxr = Intrinsic::getDeclaration(M, Int); 16959 Type *Int64Ty = Type::getInt64Ty(M->getContext()); 16960 16961 Value *Lo = Builder.CreateTrunc(Val, Int64Ty, "lo"); 16962 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 64), Int64Ty, "hi"); 16963 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext())); 16964 return Builder.CreateCall(Stxr, {Lo, Hi, Addr}); 16965 } 16966 16967 Intrinsic::ID Int = 16968 IsRelease ? Intrinsic::aarch64_stlxr : Intrinsic::aarch64_stxr; 16969 Type *Tys[] = { Addr->getType() }; 16970 Function *Stxr = Intrinsic::getDeclaration(M, Int, Tys); 16971 16972 const DataLayout &DL = M->getDataLayout(); 16973 IntegerType *IntValTy = Builder.getIntNTy(DL.getTypeSizeInBits(Val->getType())); 16974 Val = Builder.CreateBitCast(Val, IntValTy); 16975 16976 return Builder.CreateCall(Stxr, 16977 {Builder.CreateZExtOrBitCast( 16978 Val, Stxr->getFunctionType()->getParamType(0)), 16979 Addr}); 16980 } 16981 16982 bool AArch64TargetLowering::functionArgumentNeedsConsecutiveRegisters( 16983 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const { 16984 if (Ty->isArrayTy()) 16985 return true; 16986 16987 const TypeSize &TySize = Ty->getPrimitiveSizeInBits(); 16988 if (TySize.isScalable() && TySize.getKnownMinSize() > 128) 16989 return true; 16990 16991 return false; 16992 } 16993 16994 bool AArch64TargetLowering::shouldNormalizeToSelectSequence(LLVMContext &, 16995 EVT) const { 16996 return false; 16997 } 16998 16999 static Value *UseTlsOffset(IRBuilder<> &IRB, unsigned Offset) { 17000 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 17001 Function *ThreadPointerFunc = 17002 Intrinsic::getDeclaration(M, Intrinsic::thread_pointer); 17003 return IRB.CreatePointerCast( 17004 IRB.CreateConstGEP1_32(IRB.getInt8Ty(), IRB.CreateCall(ThreadPointerFunc), 17005 Offset), 17006 IRB.getInt8PtrTy()->getPointerTo(0)); 17007 } 17008 17009 Value *AArch64TargetLowering::getIRStackGuard(IRBuilder<> &IRB) const { 17010 // Android provides a fixed TLS slot for the stack cookie. See the definition 17011 // of TLS_SLOT_STACK_GUARD in 17012 // https://android.googlesource.com/platform/bionic/+/master/libc/private/bionic_tls.h 17013 if (Subtarget->isTargetAndroid()) 17014 return UseTlsOffset(IRB, 0x28); 17015 17016 // Fuchsia is similar. 17017 // <zircon/tls.h> defines ZX_TLS_STACK_GUARD_OFFSET with this value. 17018 if (Subtarget->isTargetFuchsia()) 17019 return UseTlsOffset(IRB, -0x10); 17020 17021 return TargetLowering::getIRStackGuard(IRB); 17022 } 17023 17024 void AArch64TargetLowering::insertSSPDeclarations(Module &M) const { 17025 // MSVC CRT provides functionalities for stack protection. 17026 if (Subtarget->getTargetTriple().isWindowsMSVCEnvironment()) { 17027 // MSVC CRT has a global variable holding security cookie. 17028 M.getOrInsertGlobal("__security_cookie", 17029 Type::getInt8PtrTy(M.getContext())); 17030 17031 // MSVC CRT has a function to validate security cookie. 17032 FunctionCallee SecurityCheckCookie = M.getOrInsertFunction( 17033 "__security_check_cookie", Type::getVoidTy(M.getContext()), 17034 Type::getInt8PtrTy(M.getContext())); 17035 if (Function *F = dyn_cast<Function>(SecurityCheckCookie.getCallee())) { 17036 F->setCallingConv(CallingConv::Win64); 17037 F->addAttribute(1, Attribute::AttrKind::InReg); 17038 } 17039 return; 17040 } 17041 TargetLowering::insertSSPDeclarations(M); 17042 } 17043 17044 Value *AArch64TargetLowering::getSDagStackGuard(const Module &M) const { 17045 // MSVC CRT has a global variable holding security cookie. 17046 if (Subtarget->getTargetTriple().isWindowsMSVCEnvironment()) 17047 return M.getGlobalVariable("__security_cookie"); 17048 return TargetLowering::getSDagStackGuard(M); 17049 } 17050 17051 Function *AArch64TargetLowering::getSSPStackGuardCheck(const Module &M) const { 17052 // MSVC CRT has a function to validate security cookie. 17053 if (Subtarget->getTargetTriple().isWindowsMSVCEnvironment()) 17054 return M.getFunction("__security_check_cookie"); 17055 return TargetLowering::getSSPStackGuardCheck(M); 17056 } 17057 17058 Value *AArch64TargetLowering::getSafeStackPointerLocation(IRBuilder<> &IRB) const { 17059 // Android provides a fixed TLS slot for the SafeStack pointer. See the 17060 // definition of TLS_SLOT_SAFESTACK in 17061 // https://android.googlesource.com/platform/bionic/+/master/libc/private/bionic_tls.h 17062 if (Subtarget->isTargetAndroid()) 17063 return UseTlsOffset(IRB, 0x48); 17064 17065 // Fuchsia is similar. 17066 // <zircon/tls.h> defines ZX_TLS_UNSAFE_SP_OFFSET with this value. 17067 if (Subtarget->isTargetFuchsia()) 17068 return UseTlsOffset(IRB, -0x8); 17069 17070 return TargetLowering::getSafeStackPointerLocation(IRB); 17071 } 17072 17073 bool AArch64TargetLowering::isMaskAndCmp0FoldingBeneficial( 17074 const Instruction &AndI) const { 17075 // Only sink 'and' mask to cmp use block if it is masking a single bit, since 17076 // this is likely to be fold the and/cmp/br into a single tbz instruction. It 17077 // may be beneficial to sink in other cases, but we would have to check that 17078 // the cmp would not get folded into the br to form a cbz for these to be 17079 // beneficial. 17080 ConstantInt* Mask = dyn_cast<ConstantInt>(AndI.getOperand(1)); 17081 if (!Mask) 17082 return false; 17083 return Mask->getValue().isPowerOf2(); 17084 } 17085 17086 bool AArch64TargetLowering:: 17087 shouldProduceAndByConstByHoistingConstFromShiftsLHSOfAnd( 17088 SDValue X, ConstantSDNode *XC, ConstantSDNode *CC, SDValue Y, 17089 unsigned OldShiftOpcode, unsigned NewShiftOpcode, 17090 SelectionDAG &DAG) const { 17091 // Does baseline recommend not to perform the fold by default? 17092 if (!TargetLowering::shouldProduceAndByConstByHoistingConstFromShiftsLHSOfAnd( 17093 X, XC, CC, Y, OldShiftOpcode, NewShiftOpcode, DAG)) 17094 return false; 17095 // Else, if this is a vector shift, prefer 'shl'. 17096 return X.getValueType().isScalarInteger() || NewShiftOpcode == ISD::SHL; 17097 } 17098 17099 bool AArch64TargetLowering::shouldExpandShift(SelectionDAG &DAG, 17100 SDNode *N) const { 17101 if (DAG.getMachineFunction().getFunction().hasMinSize() && 17102 !Subtarget->isTargetWindows() && !Subtarget->isTargetDarwin()) 17103 return false; 17104 return true; 17105 } 17106 17107 void AArch64TargetLowering::initializeSplitCSR(MachineBasicBlock *Entry) const { 17108 // Update IsSplitCSR in AArch64unctionInfo. 17109 AArch64FunctionInfo *AFI = Entry->getParent()->getInfo<AArch64FunctionInfo>(); 17110 AFI->setIsSplitCSR(true); 17111 } 17112 17113 void AArch64TargetLowering::insertCopiesSplitCSR( 17114 MachineBasicBlock *Entry, 17115 const SmallVectorImpl<MachineBasicBlock *> &Exits) const { 17116 const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo(); 17117 const MCPhysReg *IStart = TRI->getCalleeSavedRegsViaCopy(Entry->getParent()); 17118 if (!IStart) 17119 return; 17120 17121 const TargetInstrInfo *TII = Subtarget->getInstrInfo(); 17122 MachineRegisterInfo *MRI = &Entry->getParent()->getRegInfo(); 17123 MachineBasicBlock::iterator MBBI = Entry->begin(); 17124 for (const MCPhysReg *I = IStart; *I; ++I) { 17125 const TargetRegisterClass *RC = nullptr; 17126 if (AArch64::GPR64RegClass.contains(*I)) 17127 RC = &AArch64::GPR64RegClass; 17128 else if (AArch64::FPR64RegClass.contains(*I)) 17129 RC = &AArch64::FPR64RegClass; 17130 else 17131 llvm_unreachable("Unexpected register class in CSRsViaCopy!"); 17132 17133 Register NewVR = MRI->createVirtualRegister(RC); 17134 // Create copy from CSR to a virtual register. 17135 // FIXME: this currently does not emit CFI pseudo-instructions, it works 17136 // fine for CXX_FAST_TLS since the C++-style TLS access functions should be 17137 // nounwind. If we want to generalize this later, we may need to emit 17138 // CFI pseudo-instructions. 17139 assert(Entry->getParent()->getFunction().hasFnAttribute( 17140 Attribute::NoUnwind) && 17141 "Function should be nounwind in insertCopiesSplitCSR!"); 17142 Entry->addLiveIn(*I); 17143 BuildMI(*Entry, MBBI, DebugLoc(), TII->get(TargetOpcode::COPY), NewVR) 17144 .addReg(*I); 17145 17146 // Insert the copy-back instructions right before the terminator. 17147 for (auto *Exit : Exits) 17148 BuildMI(*Exit, Exit->getFirstTerminator(), DebugLoc(), 17149 TII->get(TargetOpcode::COPY), *I) 17150 .addReg(NewVR); 17151 } 17152 } 17153 17154 bool AArch64TargetLowering::isIntDivCheap(EVT VT, AttributeList Attr) const { 17155 // Integer division on AArch64 is expensive. However, when aggressively 17156 // optimizing for code size, we prefer to use a div instruction, as it is 17157 // usually smaller than the alternative sequence. 17158 // The exception to this is vector division. Since AArch64 doesn't have vector 17159 // integer division, leaving the division as-is is a loss even in terms of 17160 // size, because it will have to be scalarized, while the alternative code 17161 // sequence can be performed in vector form. 17162 bool OptSize = Attr.hasFnAttribute(Attribute::MinSize); 17163 return OptSize && !VT.isVector(); 17164 } 17165 17166 bool AArch64TargetLowering::preferIncOfAddToSubOfNot(EVT VT) const { 17167 // We want inc-of-add for scalars and sub-of-not for vectors. 17168 return VT.isScalarInteger(); 17169 } 17170 17171 bool AArch64TargetLowering::enableAggressiveFMAFusion(EVT VT) const { 17172 return Subtarget->hasAggressiveFMA() && VT.isFloatingPoint(); 17173 } 17174 17175 unsigned 17176 AArch64TargetLowering::getVaListSizeInBits(const DataLayout &DL) const { 17177 if (Subtarget->isTargetDarwin() || Subtarget->isTargetWindows()) 17178 return getPointerTy(DL).getSizeInBits(); 17179 17180 return 3 * getPointerTy(DL).getSizeInBits() + 2 * 32; 17181 } 17182 17183 void AArch64TargetLowering::finalizeLowering(MachineFunction &MF) const { 17184 MF.getFrameInfo().computeMaxCallFrameSize(MF); 17185 TargetLoweringBase::finalizeLowering(MF); 17186 } 17187 17188 // Unlike X86, we let frame lowering assign offsets to all catch objects. 17189 bool AArch64TargetLowering::needsFixedCatchObjects() const { 17190 return false; 17191 } 17192 17193 bool AArch64TargetLowering::shouldLocalize( 17194 const MachineInstr &MI, const TargetTransformInfo *TTI) const { 17195 switch (MI.getOpcode()) { 17196 case TargetOpcode::G_GLOBAL_VALUE: { 17197 // On Darwin, TLS global vars get selected into function calls, which 17198 // we don't want localized, as they can get moved into the middle of a 17199 // another call sequence. 17200 const GlobalValue &GV = *MI.getOperand(1).getGlobal(); 17201 if (GV.isThreadLocal() && Subtarget->isTargetMachO()) 17202 return false; 17203 break; 17204 } 17205 // If we legalized G_GLOBAL_VALUE into ADRP + G_ADD_LOW, mark both as being 17206 // localizable. 17207 case AArch64::ADRP: 17208 case AArch64::G_ADD_LOW: 17209 return true; 17210 default: 17211 break; 17212 } 17213 return TargetLoweringBase::shouldLocalize(MI, TTI); 17214 } 17215 17216 bool AArch64TargetLowering::fallBackToDAGISel(const Instruction &Inst) const { 17217 if (isa<ScalableVectorType>(Inst.getType())) 17218 return true; 17219 17220 for (unsigned i = 0; i < Inst.getNumOperands(); ++i) 17221 if (isa<ScalableVectorType>(Inst.getOperand(i)->getType())) 17222 return true; 17223 17224 if (const AllocaInst *AI = dyn_cast<AllocaInst>(&Inst)) { 17225 if (isa<ScalableVectorType>(AI->getAllocatedType())) 17226 return true; 17227 } 17228 17229 return false; 17230 } 17231 17232 // Return the largest legal scalable vector type that matches VT's element type. 17233 static EVT getContainerForFixedLengthVector(SelectionDAG &DAG, EVT VT) { 17234 assert(VT.isFixedLengthVector() && 17235 DAG.getTargetLoweringInfo().isTypeLegal(VT) && 17236 "Expected legal fixed length vector!"); 17237 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) { 17238 default: 17239 llvm_unreachable("unexpected element type for SVE container"); 17240 case MVT::i8: 17241 return EVT(MVT::nxv16i8); 17242 case MVT::i16: 17243 return EVT(MVT::nxv8i16); 17244 case MVT::i32: 17245 return EVT(MVT::nxv4i32); 17246 case MVT::i64: 17247 return EVT(MVT::nxv2i64); 17248 case MVT::f16: 17249 return EVT(MVT::nxv8f16); 17250 case MVT::f32: 17251 return EVT(MVT::nxv4f32); 17252 case MVT::f64: 17253 return EVT(MVT::nxv2f64); 17254 } 17255 } 17256 17257 // Return a PTRUE with active lanes corresponding to the extent of VT. 17258 static SDValue getPredicateForFixedLengthVector(SelectionDAG &DAG, SDLoc &DL, 17259 EVT VT) { 17260 assert(VT.isFixedLengthVector() && 17261 DAG.getTargetLoweringInfo().isTypeLegal(VT) && 17262 "Expected legal fixed length vector!"); 17263 17264 int PgPattern; 17265 switch (VT.getVectorNumElements()) { 17266 default: 17267 llvm_unreachable("unexpected element count for SVE predicate"); 17268 case 1: 17269 PgPattern = AArch64SVEPredPattern::vl1; 17270 break; 17271 case 2: 17272 PgPattern = AArch64SVEPredPattern::vl2; 17273 break; 17274 case 4: 17275 PgPattern = AArch64SVEPredPattern::vl4; 17276 break; 17277 case 8: 17278 PgPattern = AArch64SVEPredPattern::vl8; 17279 break; 17280 case 16: 17281 PgPattern = AArch64SVEPredPattern::vl16; 17282 break; 17283 case 32: 17284 PgPattern = AArch64SVEPredPattern::vl32; 17285 break; 17286 case 64: 17287 PgPattern = AArch64SVEPredPattern::vl64; 17288 break; 17289 case 128: 17290 PgPattern = AArch64SVEPredPattern::vl128; 17291 break; 17292 case 256: 17293 PgPattern = AArch64SVEPredPattern::vl256; 17294 break; 17295 } 17296 17297 // TODO: For vectors that are exactly getMaxSVEVectorSizeInBits big, we can 17298 // use AArch64SVEPredPattern::all, which can enable the use of unpredicated 17299 // variants of instructions when available. 17300 17301 MVT MaskVT; 17302 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) { 17303 default: 17304 llvm_unreachable("unexpected element type for SVE predicate"); 17305 case MVT::i8: 17306 MaskVT = MVT::nxv16i1; 17307 break; 17308 case MVT::i16: 17309 case MVT::f16: 17310 MaskVT = MVT::nxv8i1; 17311 break; 17312 case MVT::i32: 17313 case MVT::f32: 17314 MaskVT = MVT::nxv4i1; 17315 break; 17316 case MVT::i64: 17317 case MVT::f64: 17318 MaskVT = MVT::nxv2i1; 17319 break; 17320 } 17321 17322 return DAG.getNode(AArch64ISD::PTRUE, DL, MaskVT, 17323 DAG.getTargetConstant(PgPattern, DL, MVT::i64)); 17324 } 17325 17326 static SDValue getPredicateForScalableVector(SelectionDAG &DAG, SDLoc &DL, 17327 EVT VT) { 17328 assert(VT.isScalableVector() && DAG.getTargetLoweringInfo().isTypeLegal(VT) && 17329 "Expected legal scalable vector!"); 17330 auto PredTy = VT.changeVectorElementType(MVT::i1); 17331 return getPTrue(DAG, DL, PredTy, AArch64SVEPredPattern::all); 17332 } 17333 17334 static SDValue getPredicateForVector(SelectionDAG &DAG, SDLoc &DL, EVT VT) { 17335 if (VT.isFixedLengthVector()) 17336 return getPredicateForFixedLengthVector(DAG, DL, VT); 17337 17338 return getPredicateForScalableVector(DAG, DL, VT); 17339 } 17340 17341 // Grow V to consume an entire SVE register. 17342 static SDValue convertToScalableVector(SelectionDAG &DAG, EVT VT, SDValue V) { 17343 assert(VT.isScalableVector() && 17344 "Expected to convert into a scalable vector!"); 17345 assert(V.getValueType().isFixedLengthVector() && 17346 "Expected a fixed length vector operand!"); 17347 SDLoc DL(V); 17348 SDValue Zero = DAG.getConstant(0, DL, MVT::i64); 17349 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), V, Zero); 17350 } 17351 17352 // Shrink V so it's just big enough to maintain a VT's worth of data. 17353 static SDValue convertFromScalableVector(SelectionDAG &DAG, EVT VT, SDValue V) { 17354 assert(VT.isFixedLengthVector() && 17355 "Expected to convert into a fixed length vector!"); 17356 assert(V.getValueType().isScalableVector() && 17357 "Expected a scalable vector operand!"); 17358 SDLoc DL(V); 17359 SDValue Zero = DAG.getConstant(0, DL, MVT::i64); 17360 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V, Zero); 17361 } 17362 17363 // Convert all fixed length vector loads larger than NEON to masked_loads. 17364 SDValue AArch64TargetLowering::LowerFixedLengthVectorLoadToSVE( 17365 SDValue Op, SelectionDAG &DAG) const { 17366 auto Load = cast<LoadSDNode>(Op); 17367 17368 SDLoc DL(Op); 17369 EVT VT = Op.getValueType(); 17370 EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT); 17371 17372 auto NewLoad = DAG.getMaskedLoad( 17373 ContainerVT, DL, Load->getChain(), Load->getBasePtr(), Load->getOffset(), 17374 getPredicateForFixedLengthVector(DAG, DL, VT), DAG.getUNDEF(ContainerVT), 17375 Load->getMemoryVT(), Load->getMemOperand(), Load->getAddressingMode(), 17376 Load->getExtensionType()); 17377 17378 auto Result = convertFromScalableVector(DAG, VT, NewLoad); 17379 SDValue MergedValues[2] = {Result, Load->getChain()}; 17380 return DAG.getMergeValues(MergedValues, DL); 17381 } 17382 17383 static SDValue convertFixedMaskToScalableVector(SDValue Mask, 17384 SelectionDAG &DAG) { 17385 SDLoc DL(Mask); 17386 EVT InVT = Mask.getValueType(); 17387 EVT ContainerVT = getContainerForFixedLengthVector(DAG, InVT); 17388 17389 auto Op1 = convertToScalableVector(DAG, ContainerVT, Mask); 17390 auto Op2 = DAG.getConstant(0, DL, ContainerVT); 17391 auto Pg = getPredicateForFixedLengthVector(DAG, DL, InVT); 17392 17393 EVT CmpVT = Pg.getValueType(); 17394 return DAG.getNode(AArch64ISD::SETCC_MERGE_ZERO, DL, CmpVT, 17395 {Pg, Op1, Op2, DAG.getCondCode(ISD::SETNE)}); 17396 } 17397 17398 // Convert all fixed length vector loads larger than NEON to masked_loads. 17399 SDValue AArch64TargetLowering::LowerFixedLengthVectorMLoadToSVE( 17400 SDValue Op, SelectionDAG &DAG) const { 17401 auto Load = cast<MaskedLoadSDNode>(Op); 17402 17403 if (Load->getExtensionType() != ISD::LoadExtType::NON_EXTLOAD) 17404 return SDValue(); 17405 17406 SDLoc DL(Op); 17407 EVT VT = Op.getValueType(); 17408 EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT); 17409 17410 SDValue Mask = convertFixedMaskToScalableVector(Load->getMask(), DAG); 17411 17412 SDValue PassThru; 17413 bool IsPassThruZeroOrUndef = false; 17414 17415 if (Load->getPassThru()->isUndef()) { 17416 PassThru = DAG.getUNDEF(ContainerVT); 17417 IsPassThruZeroOrUndef = true; 17418 } else { 17419 if (ContainerVT.isInteger()) 17420 PassThru = DAG.getConstant(0, DL, ContainerVT); 17421 else 17422 PassThru = DAG.getConstantFP(0, DL, ContainerVT); 17423 if (isZerosVector(Load->getPassThru().getNode())) 17424 IsPassThruZeroOrUndef = true; 17425 } 17426 17427 auto NewLoad = DAG.getMaskedLoad( 17428 ContainerVT, DL, Load->getChain(), Load->getBasePtr(), Load->getOffset(), 17429 Mask, PassThru, Load->getMemoryVT(), Load->getMemOperand(), 17430 Load->getAddressingMode(), Load->getExtensionType()); 17431 17432 if (!IsPassThruZeroOrUndef) { 17433 SDValue OldPassThru = 17434 convertToScalableVector(DAG, ContainerVT, Load->getPassThru()); 17435 NewLoad = DAG.getSelect(DL, ContainerVT, Mask, NewLoad, OldPassThru); 17436 } 17437 17438 auto Result = convertFromScalableVector(DAG, VT, NewLoad); 17439 SDValue MergedValues[2] = {Result, Load->getChain()}; 17440 return DAG.getMergeValues(MergedValues, DL); 17441 } 17442 17443 // Convert all fixed length vector stores larger than NEON to masked_stores. 17444 SDValue AArch64TargetLowering::LowerFixedLengthVectorStoreToSVE( 17445 SDValue Op, SelectionDAG &DAG) const { 17446 auto Store = cast<StoreSDNode>(Op); 17447 17448 SDLoc DL(Op); 17449 EVT VT = Store->getValue().getValueType(); 17450 EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT); 17451 17452 auto NewValue = convertToScalableVector(DAG, ContainerVT, Store->getValue()); 17453 return DAG.getMaskedStore( 17454 Store->getChain(), DL, NewValue, Store->getBasePtr(), Store->getOffset(), 17455 getPredicateForFixedLengthVector(DAG, DL, VT), Store->getMemoryVT(), 17456 Store->getMemOperand(), Store->getAddressingMode(), 17457 Store->isTruncatingStore()); 17458 } 17459 17460 SDValue AArch64TargetLowering::LowerFixedLengthVectorMStoreToSVE( 17461 SDValue Op, SelectionDAG &DAG) const { 17462 auto Store = cast<MaskedStoreSDNode>(Op); 17463 17464 if (Store->isTruncatingStore()) 17465 return SDValue(); 17466 17467 SDLoc DL(Op); 17468 EVT VT = Store->getValue().getValueType(); 17469 EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT); 17470 17471 auto NewValue = convertToScalableVector(DAG, ContainerVT, Store->getValue()); 17472 SDValue Mask = convertFixedMaskToScalableVector(Store->getMask(), DAG); 17473 17474 return DAG.getMaskedStore( 17475 Store->getChain(), DL, NewValue, Store->getBasePtr(), Store->getOffset(), 17476 Mask, Store->getMemoryVT(), Store->getMemOperand(), 17477 Store->getAddressingMode(), Store->isTruncatingStore()); 17478 } 17479 17480 SDValue AArch64TargetLowering::LowerFixedLengthVectorIntDivideToSVE( 17481 SDValue Op, SelectionDAG &DAG) const { 17482 SDLoc dl(Op); 17483 EVT VT = Op.getValueType(); 17484 EVT EltVT = VT.getVectorElementType(); 17485 17486 bool Signed = Op.getOpcode() == ISD::SDIV; 17487 unsigned PredOpcode = Signed ? AArch64ISD::SDIV_PRED : AArch64ISD::UDIV_PRED; 17488 17489 // Scalable vector i32/i64 DIV is supported. 17490 if (EltVT == MVT::i32 || EltVT == MVT::i64) 17491 return LowerToPredicatedOp(Op, DAG, PredOpcode, /*OverrideNEON=*/true); 17492 17493 // Scalable vector i8/i16 DIV is not supported. Promote it to i32. 17494 EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT); 17495 EVT HalfVT = VT.getHalfNumVectorElementsVT(*DAG.getContext()); 17496 EVT FixedWidenedVT = HalfVT.widenIntegerVectorElementType(*DAG.getContext()); 17497 EVT ScalableWidenedVT = getContainerForFixedLengthVector(DAG, FixedWidenedVT); 17498 17499 // If this is not a full vector, extend, div, and truncate it. 17500 EVT WidenedVT = VT.widenIntegerVectorElementType(*DAG.getContext()); 17501 if (DAG.getTargetLoweringInfo().isTypeLegal(WidenedVT)) { 17502 unsigned ExtendOpcode = Signed ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; 17503 SDValue Op0 = DAG.getNode(ExtendOpcode, dl, WidenedVT, Op.getOperand(0)); 17504 SDValue Op1 = DAG.getNode(ExtendOpcode, dl, WidenedVT, Op.getOperand(1)); 17505 SDValue Div = DAG.getNode(Op.getOpcode(), dl, WidenedVT, Op0, Op1); 17506 return DAG.getNode(ISD::TRUNCATE, dl, VT, Div); 17507 } 17508 17509 // Convert the operands to scalable vectors. 17510 SDValue Op0 = convertToScalableVector(DAG, ContainerVT, Op.getOperand(0)); 17511 SDValue Op1 = convertToScalableVector(DAG, ContainerVT, Op.getOperand(1)); 17512 17513 // Extend the scalable operands. 17514 unsigned UnpkLo = Signed ? AArch64ISD::SUNPKLO : AArch64ISD::UUNPKLO; 17515 unsigned UnpkHi = Signed ? AArch64ISD::SUNPKHI : AArch64ISD::UUNPKHI; 17516 SDValue Op0Lo = DAG.getNode(UnpkLo, dl, ScalableWidenedVT, Op0); 17517 SDValue Op1Lo = DAG.getNode(UnpkLo, dl, ScalableWidenedVT, Op1); 17518 SDValue Op0Hi = DAG.getNode(UnpkHi, dl, ScalableWidenedVT, Op0); 17519 SDValue Op1Hi = DAG.getNode(UnpkHi, dl, ScalableWidenedVT, Op1); 17520 17521 // Convert back to fixed vectors so the DIV can be further lowered. 17522 Op0Lo = convertFromScalableVector(DAG, FixedWidenedVT, Op0Lo); 17523 Op1Lo = convertFromScalableVector(DAG, FixedWidenedVT, Op1Lo); 17524 Op0Hi = convertFromScalableVector(DAG, FixedWidenedVT, Op0Hi); 17525 Op1Hi = convertFromScalableVector(DAG, FixedWidenedVT, Op1Hi); 17526 SDValue ResultLo = DAG.getNode(Op.getOpcode(), dl, FixedWidenedVT, 17527 Op0Lo, Op1Lo); 17528 SDValue ResultHi = DAG.getNode(Op.getOpcode(), dl, FixedWidenedVT, 17529 Op0Hi, Op1Hi); 17530 17531 // Convert again to scalable vectors to truncate. 17532 ResultLo = convertToScalableVector(DAG, ScalableWidenedVT, ResultLo); 17533 ResultHi = convertToScalableVector(DAG, ScalableWidenedVT, ResultHi); 17534 SDValue ScalableResult = DAG.getNode(AArch64ISD::UZP1, dl, ContainerVT, 17535 ResultLo, ResultHi); 17536 17537 return convertFromScalableVector(DAG, VT, ScalableResult); 17538 } 17539 17540 SDValue AArch64TargetLowering::LowerFixedLengthVectorIntExtendToSVE( 17541 SDValue Op, SelectionDAG &DAG) const { 17542 EVT VT = Op.getValueType(); 17543 assert(VT.isFixedLengthVector() && "Expected fixed length vector type!"); 17544 17545 SDLoc DL(Op); 17546 SDValue Val = Op.getOperand(0); 17547 EVT ContainerVT = getContainerForFixedLengthVector(DAG, Val.getValueType()); 17548 Val = convertToScalableVector(DAG, ContainerVT, Val); 17549 17550 bool Signed = Op.getOpcode() == ISD::SIGN_EXTEND; 17551 unsigned ExtendOpc = Signed ? AArch64ISD::SUNPKLO : AArch64ISD::UUNPKLO; 17552 17553 // Repeatedly unpack Val until the result is of the desired element type. 17554 switch (ContainerVT.getSimpleVT().SimpleTy) { 17555 default: 17556 llvm_unreachable("unimplemented container type"); 17557 case MVT::nxv16i8: 17558 Val = DAG.getNode(ExtendOpc, DL, MVT::nxv8i16, Val); 17559 if (VT.getVectorElementType() == MVT::i16) 17560 break; 17561 LLVM_FALLTHROUGH; 17562 case MVT::nxv8i16: 17563 Val = DAG.getNode(ExtendOpc, DL, MVT::nxv4i32, Val); 17564 if (VT.getVectorElementType() == MVT::i32) 17565 break; 17566 LLVM_FALLTHROUGH; 17567 case MVT::nxv4i32: 17568 Val = DAG.getNode(ExtendOpc, DL, MVT::nxv2i64, Val); 17569 assert(VT.getVectorElementType() == MVT::i64 && "Unexpected element type!"); 17570 break; 17571 } 17572 17573 return convertFromScalableVector(DAG, VT, Val); 17574 } 17575 17576 SDValue AArch64TargetLowering::LowerFixedLengthVectorTruncateToSVE( 17577 SDValue Op, SelectionDAG &DAG) const { 17578 EVT VT = Op.getValueType(); 17579 assert(VT.isFixedLengthVector() && "Expected fixed length vector type!"); 17580 17581 SDLoc DL(Op); 17582 SDValue Val = Op.getOperand(0); 17583 EVT ContainerVT = getContainerForFixedLengthVector(DAG, Val.getValueType()); 17584 Val = convertToScalableVector(DAG, ContainerVT, Val); 17585 17586 // Repeatedly truncate Val until the result is of the desired element type. 17587 switch (ContainerVT.getSimpleVT().SimpleTy) { 17588 default: 17589 llvm_unreachable("unimplemented container type"); 17590 case MVT::nxv2i64: 17591 Val = DAG.getNode(ISD::BITCAST, DL, MVT::nxv4i32, Val); 17592 Val = DAG.getNode(AArch64ISD::UZP1, DL, MVT::nxv4i32, Val, Val); 17593 if (VT.getVectorElementType() == MVT::i32) 17594 break; 17595 LLVM_FALLTHROUGH; 17596 case MVT::nxv4i32: 17597 Val = DAG.getNode(ISD::BITCAST, DL, MVT::nxv8i16, Val); 17598 Val = DAG.getNode(AArch64ISD::UZP1, DL, MVT::nxv8i16, Val, Val); 17599 if (VT.getVectorElementType() == MVT::i16) 17600 break; 17601 LLVM_FALLTHROUGH; 17602 case MVT::nxv8i16: 17603 Val = DAG.getNode(ISD::BITCAST, DL, MVT::nxv16i8, Val); 17604 Val = DAG.getNode(AArch64ISD::UZP1, DL, MVT::nxv16i8, Val, Val); 17605 assert(VT.getVectorElementType() == MVT::i8 && "Unexpected element type!"); 17606 break; 17607 } 17608 17609 return convertFromScalableVector(DAG, VT, Val); 17610 } 17611 17612 SDValue AArch64TargetLowering::LowerFixedLengthExtractVectorElt( 17613 SDValue Op, SelectionDAG &DAG) const { 17614 EVT VT = Op.getValueType(); 17615 EVT InVT = Op.getOperand(0).getValueType(); 17616 assert(InVT.isFixedLengthVector() && "Expected fixed length vector type!"); 17617 17618 SDLoc DL(Op); 17619 EVT ContainerVT = getContainerForFixedLengthVector(DAG, InVT); 17620 SDValue Op0 = convertToScalableVector(DAG, ContainerVT, Op->getOperand(0)); 17621 17622 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, Op0, Op.getOperand(1)); 17623 } 17624 17625 SDValue AArch64TargetLowering::LowerFixedLengthInsertVectorElt( 17626 SDValue Op, SelectionDAG &DAG) const { 17627 EVT VT = Op.getValueType(); 17628 assert(VT.isFixedLengthVector() && "Expected fixed length vector type!"); 17629 17630 SDLoc DL(Op); 17631 EVT InVT = Op.getOperand(0).getValueType(); 17632 EVT ContainerVT = getContainerForFixedLengthVector(DAG, InVT); 17633 SDValue Op0 = convertToScalableVector(DAG, ContainerVT, Op->getOperand(0)); 17634 17635 auto ScalableRes = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ContainerVT, Op0, 17636 Op.getOperand(1), Op.getOperand(2)); 17637 17638 return convertFromScalableVector(DAG, VT, ScalableRes); 17639 } 17640 17641 // Convert vector operation 'Op' to an equivalent predicated operation whereby 17642 // the original operation's type is used to construct a suitable predicate. 17643 // NOTE: The results for inactive lanes are undefined. 17644 SDValue AArch64TargetLowering::LowerToPredicatedOp(SDValue Op, 17645 SelectionDAG &DAG, 17646 unsigned NewOp, 17647 bool OverrideNEON) const { 17648 EVT VT = Op.getValueType(); 17649 SDLoc DL(Op); 17650 auto Pg = getPredicateForVector(DAG, DL, VT); 17651 17652 if (useSVEForFixedLengthVectorVT(VT, OverrideNEON)) { 17653 EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT); 17654 17655 // Create list of operands by converting existing ones to scalable types. 17656 SmallVector<SDValue, 4> Operands = {Pg}; 17657 for (const SDValue &V : Op->op_values()) { 17658 if (isa<CondCodeSDNode>(V)) { 17659 Operands.push_back(V); 17660 continue; 17661 } 17662 17663 if (const VTSDNode *VTNode = dyn_cast<VTSDNode>(V)) { 17664 EVT VTArg = VTNode->getVT().getVectorElementType(); 17665 EVT NewVTArg = ContainerVT.changeVectorElementType(VTArg); 17666 Operands.push_back(DAG.getValueType(NewVTArg)); 17667 continue; 17668 } 17669 17670 assert(useSVEForFixedLengthVectorVT(V.getValueType(), OverrideNEON) && 17671 "Only fixed length vectors are supported!"); 17672 Operands.push_back(convertToScalableVector(DAG, ContainerVT, V)); 17673 } 17674 17675 if (isMergePassthruOpcode(NewOp)) 17676 Operands.push_back(DAG.getUNDEF(ContainerVT)); 17677 17678 auto ScalableRes = DAG.getNode(NewOp, DL, ContainerVT, Operands); 17679 return convertFromScalableVector(DAG, VT, ScalableRes); 17680 } 17681 17682 assert(VT.isScalableVector() && "Only expect to lower scalable vector op!"); 17683 17684 SmallVector<SDValue, 4> Operands = {Pg}; 17685 for (const SDValue &V : Op->op_values()) { 17686 assert((!V.getValueType().isVector() || 17687 V.getValueType().isScalableVector()) && 17688 "Only scalable vectors are supported!"); 17689 Operands.push_back(V); 17690 } 17691 17692 if (isMergePassthruOpcode(NewOp)) 17693 Operands.push_back(DAG.getUNDEF(VT)); 17694 17695 return DAG.getNode(NewOp, DL, VT, Operands); 17696 } 17697 17698 // If a fixed length vector operation has no side effects when applied to 17699 // undefined elements, we can safely use scalable vectors to perform the same 17700 // operation without needing to worry about predication. 17701 SDValue AArch64TargetLowering::LowerToScalableOp(SDValue Op, 17702 SelectionDAG &DAG) const { 17703 EVT VT = Op.getValueType(); 17704 assert(useSVEForFixedLengthVectorVT(VT) && 17705 "Only expected to lower fixed length vector operation!"); 17706 EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT); 17707 17708 // Create list of operands by converting existing ones to scalable types. 17709 SmallVector<SDValue, 4> Ops; 17710 for (const SDValue &V : Op->op_values()) { 17711 assert(!isa<VTSDNode>(V) && "Unexpected VTSDNode node!"); 17712 17713 // Pass through non-vector operands. 17714 if (!V.getValueType().isVector()) { 17715 Ops.push_back(V); 17716 continue; 17717 } 17718 17719 // "cast" fixed length vector to a scalable vector. 17720 assert(useSVEForFixedLengthVectorVT(V.getValueType()) && 17721 "Only fixed length vectors are supported!"); 17722 Ops.push_back(convertToScalableVector(DAG, ContainerVT, V)); 17723 } 17724 17725 auto ScalableRes = DAG.getNode(Op.getOpcode(), SDLoc(Op), ContainerVT, Ops); 17726 return convertFromScalableVector(DAG, VT, ScalableRes); 17727 } 17728 17729 SDValue AArch64TargetLowering::LowerVECREDUCE_SEQ_FADD(SDValue ScalarOp, 17730 SelectionDAG &DAG) const { 17731 SDLoc DL(ScalarOp); 17732 SDValue AccOp = ScalarOp.getOperand(0); 17733 SDValue VecOp = ScalarOp.getOperand(1); 17734 EVT SrcVT = VecOp.getValueType(); 17735 EVT ResVT = SrcVT.getVectorElementType(); 17736 17737 EVT ContainerVT = SrcVT; 17738 if (SrcVT.isFixedLengthVector()) { 17739 ContainerVT = getContainerForFixedLengthVector(DAG, SrcVT); 17740 VecOp = convertToScalableVector(DAG, ContainerVT, VecOp); 17741 } 17742 17743 SDValue Pg = getPredicateForVector(DAG, DL, SrcVT); 17744 SDValue Zero = DAG.getConstant(0, DL, MVT::i64); 17745 17746 // Convert operands to Scalable. 17747 AccOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ContainerVT, 17748 DAG.getUNDEF(ContainerVT), AccOp, Zero); 17749 17750 // Perform reduction. 17751 SDValue Rdx = DAG.getNode(AArch64ISD::FADDA_PRED, DL, ContainerVT, 17752 Pg, AccOp, VecOp); 17753 17754 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResVT, Rdx, Zero); 17755 } 17756 17757 SDValue AArch64TargetLowering::LowerPredReductionToSVE(SDValue ReduceOp, 17758 SelectionDAG &DAG) const { 17759 SDLoc DL(ReduceOp); 17760 SDValue Op = ReduceOp.getOperand(0); 17761 EVT OpVT = Op.getValueType(); 17762 EVT VT = ReduceOp.getValueType(); 17763 17764 if (!OpVT.isScalableVector() || OpVT.getVectorElementType() != MVT::i1) 17765 return SDValue(); 17766 17767 SDValue Pg = getPredicateForVector(DAG, DL, OpVT); 17768 17769 switch (ReduceOp.getOpcode()) { 17770 default: 17771 return SDValue(); 17772 case ISD::VECREDUCE_OR: 17773 return getPTest(DAG, VT, Pg, Op, AArch64CC::ANY_ACTIVE); 17774 case ISD::VECREDUCE_AND: { 17775 Op = DAG.getNode(ISD::XOR, DL, OpVT, Op, Pg); 17776 return getPTest(DAG, VT, Pg, Op, AArch64CC::NONE_ACTIVE); 17777 } 17778 case ISD::VECREDUCE_XOR: { 17779 SDValue ID = 17780 DAG.getTargetConstant(Intrinsic::aarch64_sve_cntp, DL, MVT::i64); 17781 SDValue Cntp = 17782 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, MVT::i64, ID, Pg, Op); 17783 return DAG.getAnyExtOrTrunc(Cntp, DL, VT); 17784 } 17785 } 17786 17787 return SDValue(); 17788 } 17789 17790 SDValue AArch64TargetLowering::LowerReductionToSVE(unsigned Opcode, 17791 SDValue ScalarOp, 17792 SelectionDAG &DAG) const { 17793 SDLoc DL(ScalarOp); 17794 SDValue VecOp = ScalarOp.getOperand(0); 17795 EVT SrcVT = VecOp.getValueType(); 17796 17797 if (useSVEForFixedLengthVectorVT(SrcVT, true)) { 17798 EVT ContainerVT = getContainerForFixedLengthVector(DAG, SrcVT); 17799 VecOp = convertToScalableVector(DAG, ContainerVT, VecOp); 17800 } 17801 17802 // UADDV always returns an i64 result. 17803 EVT ResVT = (Opcode == AArch64ISD::UADDV_PRED) ? MVT::i64 : 17804 SrcVT.getVectorElementType(); 17805 EVT RdxVT = SrcVT; 17806 if (SrcVT.isFixedLengthVector() || Opcode == AArch64ISD::UADDV_PRED) 17807 RdxVT = getPackedSVEVectorVT(ResVT); 17808 17809 SDValue Pg = getPredicateForVector(DAG, DL, SrcVT); 17810 SDValue Rdx = DAG.getNode(Opcode, DL, RdxVT, Pg, VecOp); 17811 SDValue Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResVT, 17812 Rdx, DAG.getConstant(0, DL, MVT::i64)); 17813 17814 // The VEC_REDUCE nodes expect an element size result. 17815 if (ResVT != ScalarOp.getValueType()) 17816 Res = DAG.getAnyExtOrTrunc(Res, DL, ScalarOp.getValueType()); 17817 17818 return Res; 17819 } 17820 17821 SDValue 17822 AArch64TargetLowering::LowerFixedLengthVectorSelectToSVE(SDValue Op, 17823 SelectionDAG &DAG) const { 17824 EVT VT = Op.getValueType(); 17825 SDLoc DL(Op); 17826 17827 EVT InVT = Op.getOperand(1).getValueType(); 17828 EVT ContainerVT = getContainerForFixedLengthVector(DAG, InVT); 17829 SDValue Op1 = convertToScalableVector(DAG, ContainerVT, Op->getOperand(1)); 17830 SDValue Op2 = convertToScalableVector(DAG, ContainerVT, Op->getOperand(2)); 17831 17832 // Convert the mask to a predicated (NOTE: We don't need to worry about 17833 // inactive lanes since VSELECT is safe when given undefined elements). 17834 EVT MaskVT = Op.getOperand(0).getValueType(); 17835 EVT MaskContainerVT = getContainerForFixedLengthVector(DAG, MaskVT); 17836 auto Mask = convertToScalableVector(DAG, MaskContainerVT, Op.getOperand(0)); 17837 Mask = DAG.getNode(ISD::TRUNCATE, DL, 17838 MaskContainerVT.changeVectorElementType(MVT::i1), Mask); 17839 17840 auto ScalableRes = DAG.getNode(ISD::VSELECT, DL, ContainerVT, 17841 Mask, Op1, Op2); 17842 17843 return convertFromScalableVector(DAG, VT, ScalableRes); 17844 } 17845 17846 SDValue AArch64TargetLowering::LowerFixedLengthVectorSetccToSVE( 17847 SDValue Op, SelectionDAG &DAG) const { 17848 SDLoc DL(Op); 17849 EVT InVT = Op.getOperand(0).getValueType(); 17850 EVT ContainerVT = getContainerForFixedLengthVector(DAG, InVT); 17851 17852 assert(useSVEForFixedLengthVectorVT(InVT) && 17853 "Only expected to lower fixed length vector operation!"); 17854 assert(Op.getValueType() == InVT.changeTypeToInteger() && 17855 "Expected integer result of the same bit length as the inputs!"); 17856 17857 auto Op1 = convertToScalableVector(DAG, ContainerVT, Op.getOperand(0)); 17858 auto Op2 = convertToScalableVector(DAG, ContainerVT, Op.getOperand(1)); 17859 auto Pg = getPredicateForFixedLengthVector(DAG, DL, InVT); 17860 17861 EVT CmpVT = Pg.getValueType(); 17862 auto Cmp = DAG.getNode(AArch64ISD::SETCC_MERGE_ZERO, DL, CmpVT, 17863 {Pg, Op1, Op2, Op.getOperand(2)}); 17864 17865 EVT PromoteVT = ContainerVT.changeTypeToInteger(); 17866 auto Promote = DAG.getBoolExtOrTrunc(Cmp, DL, PromoteVT, InVT); 17867 return convertFromScalableVector(DAG, Op.getValueType(), Promote); 17868 } 17869 17870 SDValue 17871 AArch64TargetLowering::LowerFixedLengthBitcastToSVE(SDValue Op, 17872 SelectionDAG &DAG) const { 17873 SDLoc DL(Op); 17874 auto SrcOp = Op.getOperand(0); 17875 EVT VT = Op.getValueType(); 17876 EVT ContainerDstVT = getContainerForFixedLengthVector(DAG, VT); 17877 EVT ContainerSrcVT = 17878 getContainerForFixedLengthVector(DAG, SrcOp.getValueType()); 17879 17880 SrcOp = convertToScalableVector(DAG, ContainerSrcVT, SrcOp); 17881 Op = DAG.getNode(ISD::BITCAST, DL, ContainerDstVT, SrcOp); 17882 return convertFromScalableVector(DAG, VT, Op); 17883 } 17884 17885 SDValue AArch64TargetLowering::getSVESafeBitCast(EVT VT, SDValue Op, 17886 SelectionDAG &DAG) const { 17887 SDLoc DL(Op); 17888 EVT InVT = Op.getValueType(); 17889 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 17890 (void)TLI; 17891 17892 assert(VT.isScalableVector() && TLI.isTypeLegal(VT) && 17893 InVT.isScalableVector() && TLI.isTypeLegal(InVT) && 17894 "Only expect to cast between legal scalable vector types!"); 17895 assert((VT.getVectorElementType() == MVT::i1) == 17896 (InVT.getVectorElementType() == MVT::i1) && 17897 "Cannot cast between data and predicate scalable vector types!"); 17898 17899 if (InVT == VT) 17900 return Op; 17901 17902 if (VT.getVectorElementType() == MVT::i1) 17903 return DAG.getNode(AArch64ISD::REINTERPRET_CAST, DL, VT, Op); 17904 17905 EVT PackedVT = getPackedSVEVectorVT(VT.getVectorElementType()); 17906 EVT PackedInVT = getPackedSVEVectorVT(InVT.getVectorElementType()); 17907 assert((VT == PackedVT || InVT == PackedInVT) && 17908 "Cannot cast between unpacked scalable vector types!"); 17909 17910 // Pack input if required. 17911 if (InVT != PackedInVT) 17912 Op = DAG.getNode(AArch64ISD::REINTERPRET_CAST, DL, PackedInVT, Op); 17913 17914 Op = DAG.getNode(ISD::BITCAST, DL, PackedVT, Op); 17915 17916 // Unpack result if required. 17917 if (VT != PackedVT) 17918 Op = DAG.getNode(AArch64ISD::REINTERPRET_CAST, DL, VT, Op); 17919 17920 return Op; 17921 } 17922 17923 bool AArch64TargetLowering::isAllActivePredicate(SDValue N) const { 17924 return ::isAllActivePredicate(N); 17925 } 17926 17927 bool AArch64TargetLowering::SimplifyDemandedBitsForTargetNode( 17928 SDValue Op, const APInt &OriginalDemandedBits, 17929 const APInt &OriginalDemandedElts, KnownBits &Known, TargetLoweringOpt &TLO, 17930 unsigned Depth) const { 17931 17932 unsigned Opc = Op.getOpcode(); 17933 switch (Opc) { 17934 case AArch64ISD::VSHL: { 17935 // Match (VSHL (VLSHR Val X) X) 17936 SDValue ShiftL = Op; 17937 SDValue ShiftR = Op->getOperand(0); 17938 if (ShiftR->getOpcode() != AArch64ISD::VLSHR) 17939 return false; 17940 17941 if (!ShiftL.hasOneUse() || !ShiftR.hasOneUse()) 17942 return false; 17943 17944 unsigned ShiftLBits = ShiftL->getConstantOperandVal(1); 17945 unsigned ShiftRBits = ShiftR->getConstantOperandVal(1); 17946 17947 // Other cases can be handled as well, but this is not 17948 // implemented. 17949 if (ShiftRBits != ShiftLBits) 17950 return false; 17951 17952 unsigned ScalarSize = Op.getScalarValueSizeInBits(); 17953 assert(ScalarSize > ShiftLBits && "Invalid shift imm"); 17954 17955 APInt ZeroBits = APInt::getLowBitsSet(ScalarSize, ShiftLBits); 17956 APInt UnusedBits = ~OriginalDemandedBits; 17957 17958 if ((ZeroBits & UnusedBits) != ZeroBits) 17959 return false; 17960 17961 // All bits that are zeroed by (VSHL (VLSHR Val X) X) are not 17962 // used - simplify to just Val. 17963 return TLO.CombineTo(Op, ShiftR->getOperand(0)); 17964 } 17965 } 17966 17967 return TargetLowering::SimplifyDemandedBitsForTargetNode( 17968 Op, OriginalDemandedBits, OriginalDemandedElts, Known, TLO, Depth); 17969 } 17970